lite node

exo theme
EXO iOS app
2026-02-19 22:07:34 +05:00 · 2026-02-19 04:21:58 +05:00 · 2026-02-18 06:40:07 +05:00 · 2026-02-17 14:31:47 +00:00 · 2026-02-17 11:48:28 +00:00 · 2026-02-17 11:43:06 +00:00
115 changed files with 7234 additions and 1721 deletions
@@ -396,7 +396,7 @@ jobs:
          path: output/EXO-${{ env.RELEASE_VERSION }}.dmg

      - name: Upload to S3
-        if: env.SPARKLE_S3_BUCKET != '' && github.ref_type == 'tag'
+        if: env.SPARKLE_S3_BUCKET != ''
        env:
          AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
          AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
@@ -412,6 +412,12 @@ jobs:
            PREFIX="${PREFIX}/"
          fi
          DMG_NAME="EXO-${RELEASE_VERSION}.dmg"
+
+          if [[ "${{ github.ref_type }}" != "tag" ]]; then
+            aws s3 cp "$DMG_NAME" "s3://${SPARKLE_S3_BUCKET}/${PREFIX}EXO-${GITHUB_SHA}.dmg"
+            exit 0
+          fi
+
          aws s3 cp "$DMG_NAME" "s3://${SPARKLE_S3_BUCKET}/${PREFIX}${DMG_NAME}"
          if [[ "$IS_ALPHA" != "true" ]]; then
            aws s3 cp "$DMG_NAME" "s3://${SPARKLE_S3_BUCKET}/${PREFIX}EXO-latest.dmg"
@@ -8,33 +8,6 @@ on:
      - main

 jobs:
-  typecheck:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-        with:
-          lfs: false
-
-      - uses: cachix/install-nix-action@v31
-        with:
-          nix_path: nixpkgs=channel:nixos-unstable
-
-      - uses: cachix/cachix-action@v14
-        name: Configure Cachix
-        with:
-          name: exo
-          authToken: "${{ secrets.CACHIX_AUTH_TOKEN }}"
-
-      - name: Load nix develop environment
-        run: nix run github:nicknovitski/nix-develop/v1
-
-      - name: Sync dependencies
-        run: uv sync --all-packages
-
-      - name: Run type checker
-        run: uv run basedpyright --project pyproject.toml
-
  nix:
    name: Build and check (${{ matrix.system }})
    runs-on: ${{ matrix.runner }}
@@ -0,0 +1,151 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs
+from .switch_layers import SwitchGLU
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    vocab_size: int
+    num_attention_heads: int
+    num_attention_groups: int
+    head_dim: int
+    intermediate_size: int
+    rms_norm_eps: float
+    rope_theta: float
+    rope_scaling: Optional[Dict[str, Any]]
+    max_position_embeddings: int
+    sliding_window: int
+    layer_types: Optional[List[str]]
+    yarn_only_types: Optional[List[str]]
+    partial_rotary_factors: Optional[List[float]]
+    attention_other_setting: Optional[Dict[str, Any]]
+    use_head_wise_attn_gate: bool
+    moe_num_experts: int
+    moe_top_k: int
+    moe_intermediate_size: int
+    share_expert_dim: int
+    moe_layers_enum: Optional[str]
+    moe_router_scaling_factor: float
+    norm_expert_weight: bool
+    swiglu_limits: Optional[List[float]]
+    swiglu_limits_shared: Optional[List[float]]
+    tie_word_embeddings: bool
+
+class Step3p5MLP(nn.Module):
+    hidden_size: int
+    intermediate_size: int
+    gate_proj: nn.Linear
+    up_proj: nn.Linear
+    down_proj: nn.Linear
+    limit: Optional[float]
+
+    def __init__(
+        self, args: ModelArgs, intermediate_size: int, swiglu_limit: float = 0
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class Step3p5MoEGate(nn.Module):
+    top_k: int
+    n_routed_experts: int
+    routed_scaling_factor: float
+    norm_topk_prob: bool
+    gate: nn.Linear
+    router_bias: mx.array
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(self, x: mx.array) -> tuple[mx.array, mx.array]: ...
+
+class Step3p5MoE(nn.Module):
+    gate: Step3p5MoEGate
+    switch_mlp: SwitchGLU
+    share_expert: Step3p5MLP
+    sharding_group: Optional[mx.distributed.Group]
+
+    def __init__(self, args: ModelArgs, layer_idx: int) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class Step3p5Attention(nn.Module):
+    is_sliding: bool
+    num_heads: int
+    num_kv_heads: int
+    head_dim: int
+    scale: float
+    q_proj: nn.Linear
+    k_proj: nn.Linear
+    v_proj: nn.Linear
+    o_proj: nn.Linear
+    q_norm: nn.Module
+    k_norm: nn.Module
+    use_head_wise_attn_gate: bool
+    g_proj: nn.Linear
+    rope: nn.Module
+
+    def __init__(self, args: ModelArgs, layer_idx: int) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class Step3p5DecoderLayer(nn.Module):
+    self_attn: Step3p5Attention
+    is_sliding: bool
+    is_moe_layer: bool
+    mlp: Step3p5MLP | Step3p5MoE
+    input_layernorm: nn.Module
+    post_attention_layernorm: nn.Module
+
+    def __init__(self, args: ModelArgs, layer_idx: int) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class Step3p5Model(nn.Module):
+    args: ModelArgs
+    vocab_size: int
+    num_layers: int
+    embed_tokens: nn.Embedding
+    layers: list[Step3p5DecoderLayer]
+    norm: nn.Module
+    _swa_idx: Optional[int]
+    _full_idx: Optional[int]
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        cache: Optional[List[Any]] = None,
+    ) -> mx.array: ...
+
+class Model(nn.Module):
+    args: ModelArgs
+    model_type: str
+    model: Step3p5Model
+    lm_head: nn.Linear
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[List[Any]] = None,
+    ) -> mx.array: ...
+    def sanitize(self, weights: dict[str, Any]) -> dict[str, Any]: ...
+    def shard(self, group: Optional[mx.distributed.Group] = None) -> None: ...
+    @property
+    def layers(self) -> list[Step3p5DecoderLayer]: ...
+    def make_cache(self) -> list[Any]: ...
+    @property
+    def cast_predicate(self) -> Any: ...
+    @property
+    def quant_predicate(self) -> Any: ...
@@ -119,3 +119,78 @@ From .cursorrules:
 ## Testing

 Tests use pytest-asyncio with `asyncio_mode = "auto"`. Tests are in `tests/` subdirectories alongside the code they test. The `EXO_TESTS=1` env var is set during tests.
+
+## Dashboard UI Testing & Screenshots
+
+### Building and Running the Dashboard
+```bash
+# Build the dashboard (must be done before running exo)
+cd dashboard && npm install && npm run build && cd ..
+
+# Start exo (serves the dashboard at http://localhost:52415)
+uv run exo &
+sleep 8  # Wait for server to start
+```
+
+### Taking Headless Screenshots with Playwright
+Use Playwright with headless Chromium for programmatic screenshots — no manual browser interaction needed.
+
+**Setup (one-time):**
+```bash
+npx --yes playwright install chromium
+cd /tmp && npm init -y && npm install playwright
+```
+
+**Taking screenshots:**
+```javascript
+// Run from /tmp where playwright is installed: cd /tmp && node -e "..."
+const { chromium } = require('playwright');
+(async () => {
+  const browser = await chromium.launch({ headless: true });
+  const page = await browser.newPage({ viewport: { width: 1280, height: 800 } });
+  await page.goto('http://localhost:52415', { waitUntil: 'networkidle' });
+  await page.waitForTimeout(2000);
+
+  // Inject test data into localStorage if needed (e.g., recent models)
+  await page.evaluate(() => {
+    localStorage.setItem('exo-recent-models', JSON.stringify([
+      { modelId: 'mlx-community/Qwen3-30B-A3B-4bit', launchedAt: Date.now() },
+    ]));
+  });
+  await page.reload({ waitUntil: 'networkidle' });
+  await page.waitForTimeout(2000);
+
+  // Interact with UI elements
+  await page.locator('text=SELECT MODEL').click();
+  await page.waitForTimeout(1000);
+
+  // Take screenshot
+  await page.screenshot({ path: '/tmp/screenshot.png', fullPage: false });
+  await browser.close();
+})();
+```
+
+### Uploading Images to GitHub PRs
+GitHub's API doesn't support direct image upload for PR comments. Workaround:
+
+1. **Commit images to the branch** (temporarily):
+   ```bash
+   cp /tmp/screenshot.png .
+   git add screenshot.png
+   git commit -m "temp: add screenshots for PR"
+   git push origin <branch>
+   COMMIT_SHA=$(git rev-parse HEAD)
+   ```
+
+2. **Post PR comment** referencing the raw image URL (uses permanent commit SHA so images survive deletion):
+   ```bash
+   gh pr comment <PR_NUMBER> --body "![Screenshot](https://raw.githubusercontent.com/exo-explore/exo/${COMMIT_SHA}/screenshot.png)"
+   ```
+
+3. **Remove the images** from the branch:
+   ```bash
+   git rm screenshot.png
+   git commit -m "chore: remove temporary screenshot files"
+   git push origin <branch>
+   ```
+   The images still render in the PR comment because they reference the permanent commit SHA.
@@ -353,31 +353,6 @@ dependencies = [
 "generic-array",
 ]

-[[package]]
-name = "bon"
-version = "3.8.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ebeb9aaf9329dff6ceb65c689ca3db33dbf15f324909c60e4e5eef5701ce31b1"
-dependencies = [
- "bon-macros",
- "rustversion",
-]
-
-[[package]]
-name = "bon-macros"
-version = "3.8.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "77e9d642a7e3a318e37c2c9427b5a6a48aa1ad55dcd986f3034ab2239045a645"
-dependencies = [
- "darling",
- "ident_case",
- "prettyplease",
- "proc-macro2",
- "quote",
- "rustversion",
- "syn 2.0.111",
-]
-
 [[package]]
 name = "bs58"
 version = "0.5.1"
@@ -700,41 +675,6 @@ dependencies = [
 "syn 2.0.111",
 ]

-[[package]]
-name = "darling"
-version = "0.21.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9cdf337090841a411e2a7f3deb9187445851f91b309c0c0a29e05f74a00a48c0"
-dependencies = [
- "darling_core",
- "darling_macro",
-]
-
-[[package]]
-name = "darling_core"
-version = "0.21.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1247195ecd7e3c85f83c8d2a366e4210d588e802133e1e355180a9870b517ea4"
-dependencies = [
- "fnv",
- "ident_case",
- "proc-macro2",
- "quote",
- "strsim",
- "syn 2.0.111",
-]
-
-[[package]]
-name = "darling_macro"
-version = "0.21.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d38308df82d1080de0afee5d069fa14b0326a88c14f15c5ccda35b4a6c414c81"
-dependencies = [
- "darling_core",
- "quote",
- "syn 2.0.111",
-]
-
 [[package]]
 name = "data-encoding"
 version = "2.9.0"
@@ -1625,12 +1565,6 @@ dependencies = [
 "zerovec",
 ]

-[[package]]
-name = "ident_case"
-version = "1.0.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b9e0384b61958566e926dc50660321d12159025e767c18e043daf26b70104c39"
-
 [[package]]
 name = "idna"
 version = "1.1.0"
@@ -1745,15 +1679,6 @@ dependencies = [
 "generic-array",
 ]

-[[package]]
-name = "internment"
-version = "0.8.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "636d4b0f6a39fd684effe2a73f5310df16a3fa7954c26d36833e98f44d1977a2"
-dependencies = [
- "hashbrown 0.15.5",
-]
-
 [[package]]
 name = "inventory"
 version = "0.3.21"
@@ -3307,16 +3232,6 @@ dependencies = [
 "zerocopy",
 ]

-[[package]]
-name = "prettyplease"
-version = "0.2.37"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "479ca8adacdd7ce8f1fb39ce9ecccbfe93a3f1344b3d0d97f20bc0196208f62b"
-dependencies = [
- "proc-macro2",
- "syn 2.0.111",
-]
-
 [[package]]
 name = "primeorder"
 version = "0.13.6"
@@ -3740,12 +3655,6 @@ dependencies = [
 "yasna",
 ]

-[[package]]
-name = "recursion"
-version = "0.5.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9dba2197bf7b1d87b4dd460c195f4edeb45a94e82e8054f8d5f317c1f0e93ca1"
-
 [[package]]
 name = "redox_syscall"
 version = "0.5.18"
@@ -4807,15 +4716,6 @@ checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"
 [[package]]
 name = "util"
 version = "0.0.1"
-dependencies = [
- "bon",
- "derive_more",
- "extend",
- "internment",
- "once_cell",
- "recursion",
- "thiserror 2.0.17",
-]

 [[package]]
 name = "uuid"
@@ -5,21 +5,21 @@
 [X] Fetching download status of all models on start
 [X] Deduplication of tasks in plan_step.
 [X] resolve_allow_patterns should just be wildcard now.
-[] no mx_barrier in genreate.py mlx_generate at the end.
+[X] no mx_barrier in genreate.py mlx_generate at the end.
 [] cache assertion not needed in auto_parallel.py PipelineLastLayer.
-[] GPTOSS support dropped in auto_parallel.py.
-[] sharding changed "all-to-sharded" became _all_to_sharded in auto_parallel.py.
-[] same as above with "sharded-to-all" became _sharded_to_all in auto_parallel.py.
-[] Dropped support for Ministral3Model, DeepseekV32Model, Glm4MoeModel, Qwen3NextModel, GptOssMode in auto_parallel.py.
+[X] GPTOSS support dropped in auto_parallel.py.
+[X] sharding changed "all-to-sharded" became _all_to_sharded in auto_parallel.py.
+[X] same as above with "sharded-to-all" became _sharded_to_all in auto_parallel.py.
+[X] Dropped support for Ministral3Model, DeepseekV32Model, Glm4MoeModel, Qwen3NextModel, GptOssMode in auto_parallel.py.
 [] Dropped prefill/decode code in auto_parallel.py and utils_mlx.py.
 [X] KV_CACHE_BITS should be None to disable quantized KV cache.
-[] Dropped _set_nofile_limit in utils_mlx.py.
-[] We have group optional in load_mlx_items in utils_mlx.py.
-[] Dropped add_missing_chat_templates for GptOss in load_mlx_items in utils_mlx.py.
-[] Dropped model.make_cache in make_kv_cache in utils_mlx.py.
+[X] Dropped _set_nofile_limit in utils_mlx.py.
+[X] We have group optional in load_mlx_items in utils_mlx.py.
+[X] Dropped add_missing_chat_templates for GptOss in load_mlx_items in utils_mlx.py.
+[X] Dropped model.make_cache in make_kv_cache in utils_mlx.py.
 [X] We put cache limit back in utils_mlx.py.
-[] topology.py remove_node removes the connections after checking if node is is in self._node_id_to_rx_id_map. on beta_1 it checks after, so would remove stale connections I guess?
-[] Missing Glm 4.7 model cards (this isn't ready yet but should be picked up, probably create an issue... the blocker is transforemrs version doesn't support the tokenizer for Glm 4.7. rc-1 does but we can't upgrade as it breaks other things.)
+[X] topology.py remove_node removes the connections after checking if node is is in self._node_id_to_rx_id_map. on beta_1 it checks after, so would remove stale connections I guess?
+[X] Missing Glm 4.7 model cards (this isn't ready yet but should be picked up, probably create an issue... the blocker is transforemrs version doesn't support the tokenizer for Glm 4.7. rc-1 does but we can't upgrade as it breaks other things.)
 [] try-except in _command_processor only excepts ValueError. This was silently failing leading to un-debuggable errors (we had a KeyError that was happening ). Changed this to catch Exception instead of ValueError. See exo-v2 89ae38405e0052e3c22405daf094b065878aa873 and fb99fea69b5a39017efc90c5dad0072e677455f0.
 [X] In placement.py, place_instance no longer looks at model_meta.supports_tensor and check if this tensor parallel number of nodes is supported by the model's tensor dimensions.
 [X] In placement.py, place_instanec, we no longer have the special case to exclude DeepSeek v3.1 pipeline parallel (it doesn't work).
@@ -1,28 +1,14 @@
-2. Currently a lot of requests from the API are timing out, but we still process those requests internally. If an API request times out, we should cancel all corresponding tasks to that API request (why process a request with nobody listening).
 3. Task cancellation. When API http request gets cancelled, it should cancel corresponding task.
 4. I'd like to see profiled network latency / bandwidth.
 5. I'd like to see how much bandwidth each link is using.
-6. We should handle the case where one machine doesn't have the model downloaded and then other machines are waiting on it. In this case we get loads of timeout errors because the others are waiting for the one that needs to download the model.
 7. Solve the problem of in continuous batching when a new prompt comes in, it will block decode of the current batch until the prefill is complete.
 8. We want people to be able to copy models over to a new device without ever connecting EXO to the internet. Right now EXO require internet connection once to cache some files to check if a download is complete. Instead, we should simply check if there is a non-empty model folder locally with no .partial files. This indicates it's a fully downloaded model that can be loaded.
-10. More granular control over how to deploy instances.
-12. Nix is great but installing it is a pain and we have ended up in a lot of cases having PATH issues or installation issues. For example, after rebooting mike it seemed to no longer have a nix installation and needed reinstalling. It has a bunch of broken symlinks left over from nix that caused ssh to fail, making it even harder to debug. We need consistent environments (perhaps MDM) so we can guarantee nix is installed properly on each machine.
 13. Memory pressure instead of memory used.
 14. Show the type of each connection (TB5, Ethernet, etc.) in the UI. Refer to old exo: https://github.com/exo-explore/exo/blob/56f783b38dc6b08ce606b07a5386dc40dae00330/exo/helpers.py#L251
 15. Prioritise certain connection types (or by latency). TB5 > Ethernet > WiFi. Refer to old exo: https://github.com/exo-explore/exo/blob/56f783b38dc6b08ce606b07a5386dc40dae00330/exo/helpers.py#L251
 16. Dynamically switch to higher priority connection when it becomes available. Probably bring back InstanceReplacedAtomically.
 17. Faster model loads by streaming model from other devices in cluster.
 18. Add support for specifying the type of network connection to use in a test. Depends on 15/16.
-20. Add chat completion cancellations (e.g OpenWebUI has something for cancelling an ongoing request).
-23. Do we need cache_limit? We went back and forth on that a lot because we thought it might be causing issues. One problem is it sets it relative to model size. So if you have multiple models loaded in it will take the most recent model size for the cache_limit. This is problematic if you launch DeepSeek -> Llama for example.
-24. further openai/lmstudio api compatibility
 25. Rethink retry logic
-26. Task cancellation. When API http request gets cancelled, it should cancel corresponding task.
 27. Log cleanup - per-module log filters and default to DEBUG log levels
 28. Validate RDMA connections with ibv_devinfo in the info gatherer
-
-Potential refactors:
-
-2. Topology can be simplified
-
-Random errors we've run into:
@@ -0,0 +1,628 @@
+// !$*UTF8*$!
+{
+	archiveVersion = 1;
+	classes = {
+	};
+	objectVersion = 77;
+	objects = {
+
+/* Begin PBXBuildFile section */
+		E09D17522F44F359009C51A3 /* MLXLLM in Frameworks */ = {isa = PBXBuildFile; productRef = E09D17512F44F359009C51A3 /* MLXLLM */; };
+		E09D17542F44F359009C51A3 /* MLXLMCommon in Frameworks */ = {isa = PBXBuildFile; productRef = E09D17532F44F359009C51A3 /* MLXLMCommon */; };
+/* End PBXBuildFile section */
+
+/* Begin PBXContainerItemProxy section */
+		E09D167D2F44CA20009C51A3 /* PBXContainerItemProxy */ = {
+			isa = PBXContainerItemProxy;
+			containerPortal = E09D16672F44CA1E009C51A3 /* Project object */;
+			proxyType = 1;
+			remoteGlobalIDString = E09D166E2F44CA1E009C51A3;
+			remoteInfo = "EXO-iOS";
+		};
+		E09D16872F44CA20009C51A3 /* PBXContainerItemProxy */ = {
+			isa = PBXContainerItemProxy;
+			containerPortal = E09D16672F44CA1E009C51A3 /* Project object */;
+			proxyType = 1;
+			remoteGlobalIDString = E09D166E2F44CA1E009C51A3;
+			remoteInfo = "EXO-iOS";
+		};
+/* End PBXContainerItemProxy section */
+
+/* Begin PBXFileReference section */
+		E09D166F2F44CA1E009C51A3 /* EXO-iOS.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = "EXO-iOS.app"; sourceTree = BUILT_PRODUCTS_DIR; };
+		E09D167C2F44CA20009C51A3 /* EXO-iOSTests.xctest */ = {isa = PBXFileReference; explicitFileType = wrapper.cfbundle; includeInIndex = 0; path = "EXO-iOSTests.xctest"; sourceTree = BUILT_PRODUCTS_DIR; };
+		E09D16862F44CA20009C51A3 /* EXO-iOSUITests.xctest */ = {isa = PBXFileReference; explicitFileType = wrapper.cfbundle; includeInIndex = 0; path = "EXO-iOSUITests.xctest"; sourceTree = BUILT_PRODUCTS_DIR; };
+/* End PBXFileReference section */
+
+/* Begin PBXFileSystemSynchronizedBuildFileExceptionSet section */
+		E09D169A2F44CA20009C51A3 /* Exceptions for "EXO-iOS" folder in "EXO-iOS" target */ = {
+			isa = PBXFileSystemSynchronizedBuildFileExceptionSet;
+			membershipExceptions = (
+				Info.plist,
+			);
+			target = E09D166E2F44CA1E009C51A3 /* EXO-iOS */;
+		};
+/* End PBXFileSystemSynchronizedBuildFileExceptionSet section */
+
+/* Begin PBXFileSystemSynchronizedRootGroup section */
+		E09D16712F44CA1E009C51A3 /* EXO-iOS */ = {
+			isa = PBXFileSystemSynchronizedRootGroup;
+			exceptions = (
+				E09D169A2F44CA20009C51A3 /* Exceptions for "EXO-iOS" folder in "EXO-iOS" target */,
+			);
+			path = "EXO-iOS";
+			sourceTree = "<group>";
+		};
+		E09D167F2F44CA20009C51A3 /* EXO-iOSTests */ = {
+			isa = PBXFileSystemSynchronizedRootGroup;
+			path = "EXO-iOSTests";
+			sourceTree = "<group>";
+		};
+		E09D16892F44CA20009C51A3 /* EXO-iOSUITests */ = {
+			isa = PBXFileSystemSynchronizedRootGroup;
+			path = "EXO-iOSUITests";
+			sourceTree = "<group>";
+		};
+/* End PBXFileSystemSynchronizedRootGroup section */
+
+/* Begin PBXFrameworksBuildPhase section */
+		E09D166C2F44CA1E009C51A3 /* Frameworks */ = {
+			isa = PBXFrameworksBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+				E09D17542F44F359009C51A3 /* MLXLMCommon in Frameworks */,
+				E09D17522F44F359009C51A3 /* MLXLLM in Frameworks */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+		E09D16792F44CA20009C51A3 /* Frameworks */ = {
+			isa = PBXFrameworksBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+		E09D16832F44CA20009C51A3 /* Frameworks */ = {
+			isa = PBXFrameworksBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXFrameworksBuildPhase section */
+
+/* Begin PBXGroup section */
+		E09D16662F44CA1E009C51A3 = {
+			isa = PBXGroup;
+			children = (
+				E09D16712F44CA1E009C51A3 /* EXO-iOS */,
+				E09D167F2F44CA20009C51A3 /* EXO-iOSTests */,
+				E09D16892F44CA20009C51A3 /* EXO-iOSUITests */,
+				E09D16702F44CA1E009C51A3 /* Products */,
+			);
+			sourceTree = "<group>";
+		};
+		E09D16702F44CA1E009C51A3 /* Products */ = {
+			isa = PBXGroup;
+			children = (
+				E09D166F2F44CA1E009C51A3 /* EXO-iOS.app */,
+				E09D167C2F44CA20009C51A3 /* EXO-iOSTests.xctest */,
+				E09D16862F44CA20009C51A3 /* EXO-iOSUITests.xctest */,
+			);
+			name = Products;
+			sourceTree = "<group>";
+		};
+/* End PBXGroup section */
+
+/* Begin PBXNativeTarget section */
+		E09D166E2F44CA1E009C51A3 /* EXO-iOS */ = {
+			isa = PBXNativeTarget;
+			buildConfigurationList = E09D16902F44CA20009C51A3 /* Build configuration list for PBXNativeTarget "EXO-iOS" */;
+			buildPhases = (
+				E09D166B2F44CA1E009C51A3 /* Sources */,
+				E09D166C2F44CA1E009C51A3 /* Frameworks */,
+				E09D166D2F44CA1E009C51A3 /* Resources */,
+			);
+			buildRules = (
+			);
+			dependencies = (
+			);
+			fileSystemSynchronizedGroups = (
+				E09D16712F44CA1E009C51A3 /* EXO-iOS */,
+			);
+			name = "EXO-iOS";
+			packageProductDependencies = (
+				E09D17512F44F359009C51A3 /* MLXLLM */,
+				E09D17532F44F359009C51A3 /* MLXLMCommon */,
+			);
+			productName = "EXO-iOS";
+			productReference = E09D166F2F44CA1E009C51A3 /* EXO-iOS.app */;
+			productType = "com.apple.product-type.application";
+		};
+		E09D167B2F44CA20009C51A3 /* EXO-iOSTests */ = {
+			isa = PBXNativeTarget;
+			buildConfigurationList = E09D16932F44CA20009C51A3 /* Build configuration list for PBXNativeTarget "EXO-iOSTests" */;
+			buildPhases = (
+				E09D16782F44CA20009C51A3 /* Sources */,
+				E09D16792F44CA20009C51A3 /* Frameworks */,
+				E09D167A2F44CA20009C51A3 /* Resources */,
+			);
+			buildRules = (
+			);
+			dependencies = (
+				E09D167E2F44CA20009C51A3 /* PBXTargetDependency */,
+			);
+			fileSystemSynchronizedGroups = (
+				E09D167F2F44CA20009C51A3 /* EXO-iOSTests */,
+			);
+			name = "EXO-iOSTests";
+			packageProductDependencies = (
+			);
+			productName = "EXO-iOSTests";
+			productReference = E09D167C2F44CA20009C51A3 /* EXO-iOSTests.xctest */;
+			productType = "com.apple.product-type.bundle.unit-test";
+		};
+		E09D16852F44CA20009C51A3 /* EXO-iOSUITests */ = {
+			isa = PBXNativeTarget;
+			buildConfigurationList = E09D16962F44CA20009C51A3 /* Build configuration list for PBXNativeTarget "EXO-iOSUITests" */;
+			buildPhases = (
+				E09D16822F44CA20009C51A3 /* Sources */,
+				E09D16832F44CA20009C51A3 /* Frameworks */,
+				E09D16842F44CA20009C51A3 /* Resources */,
+			);
+			buildRules = (
+			);
+			dependencies = (
+				E09D16882F44CA20009C51A3 /* PBXTargetDependency */,
+			);
+			fileSystemSynchronizedGroups = (
+				E09D16892F44CA20009C51A3 /* EXO-iOSUITests */,
+			);
+			name = "EXO-iOSUITests";
+			packageProductDependencies = (
+			);
+			productName = "EXO-iOSUITests";
+			productReference = E09D16862F44CA20009C51A3 /* EXO-iOSUITests.xctest */;
+			productType = "com.apple.product-type.bundle.ui-testing";
+		};
+/* End PBXNativeTarget section */
+
+/* Begin PBXProject section */
+		E09D16672F44CA1E009C51A3 /* Project object */ = {
+			isa = PBXProject;
+			attributes = {
+				BuildIndependentTargetsInParallel = 1;
+				LastSwiftUpdateCheck = 2620;
+				LastUpgradeCheck = 2620;
+				TargetAttributes = {
+					E09D166E2F44CA1E009C51A3 = {
+						CreatedOnToolsVersion = 26.2;
+					};
+					E09D167B2F44CA20009C51A3 = {
+						CreatedOnToolsVersion = 26.2;
+						TestTargetID = E09D166E2F44CA1E009C51A3;
+					};
+					E09D16852F44CA20009C51A3 = {
+						CreatedOnToolsVersion = 26.2;
+						TestTargetID = E09D166E2F44CA1E009C51A3;
+					};
+				};
+			};
+			buildConfigurationList = E09D166A2F44CA1E009C51A3 /* Build configuration list for PBXProject "EXO-iOS" */;
+			developmentRegion = en;
+			hasScannedForEncodings = 0;
+			knownRegions = (
+				en,
+				Base,
+			);
+			mainGroup = E09D16662F44CA1E009C51A3;
+			minimizedProjectReferenceProxies = 1;
+			packageReferences = (
+				E09D17502F44F359009C51A3 /* XCRemoteSwiftPackageReference "mlx-swift-lm" */,
+			);
+			preferredProjectObjectVersion = 77;
+			productRefGroup = E09D16702F44CA1E009C51A3 /* Products */;
+			projectDirPath = "";
+			projectRoot = "";
+			targets = (
+				E09D166E2F44CA1E009C51A3 /* EXO-iOS */,
+				E09D167B2F44CA20009C51A3 /* EXO-iOSTests */,
+				E09D16852F44CA20009C51A3 /* EXO-iOSUITests */,
+			);
+		};
+/* End PBXProject section */
+
+/* Begin PBXResourcesBuildPhase section */
+		E09D166D2F44CA1E009C51A3 /* Resources */ = {
+			isa = PBXResourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+		E09D167A2F44CA20009C51A3 /* Resources */ = {
+			isa = PBXResourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+		E09D16842F44CA20009C51A3 /* Resources */ = {
+			isa = PBXResourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXResourcesBuildPhase section */
+
+/* Begin PBXSourcesBuildPhase section */
+		E09D166B2F44CA1E009C51A3 /* Sources */ = {
+			isa = PBXSourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+		E09D16782F44CA20009C51A3 /* Sources */ = {
+			isa = PBXSourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+		E09D16822F44CA20009C51A3 /* Sources */ = {
+			isa = PBXSourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXSourcesBuildPhase section */
+
+/* Begin PBXTargetDependency section */
+		E09D167E2F44CA20009C51A3 /* PBXTargetDependency */ = {
+			isa = PBXTargetDependency;
+			target = E09D166E2F44CA1E009C51A3 /* EXO-iOS */;
+			targetProxy = E09D167D2F44CA20009C51A3 /* PBXContainerItemProxy */;
+		};
+		E09D16882F44CA20009C51A3 /* PBXTargetDependency */ = {
+			isa = PBXTargetDependency;
+			target = E09D166E2F44CA1E009C51A3 /* EXO-iOS */;
+			targetProxy = E09D16872F44CA20009C51A3 /* PBXContainerItemProxy */;
+		};
+/* End PBXTargetDependency section */
+
+/* Begin XCBuildConfiguration section */
+		E09D168E2F44CA20009C51A3 /* Debug */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ALWAYS_SEARCH_USER_PATHS = NO;
+				ASSETCATALOG_COMPILER_GENERATE_SWIFT_ASSET_SYMBOL_EXTENSIONS = YES;
+				CLANG_ANALYZER_NONNULL = YES;
+				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
+				CLANG_CXX_LANGUAGE_STANDARD = "gnu++20";
+				CLANG_ENABLE_MODULES = YES;
+				CLANG_ENABLE_OBJC_ARC = YES;
+				CLANG_ENABLE_OBJC_WEAK = YES;
+				CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
+				CLANG_WARN_BOOL_CONVERSION = YES;
+				CLANG_WARN_COMMA = YES;
+				CLANG_WARN_CONSTANT_CONVERSION = YES;
+				CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
+				CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
+				CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
+				CLANG_WARN_EMPTY_BODY = YES;
+				CLANG_WARN_ENUM_CONVERSION = YES;
+				CLANG_WARN_INFINITE_RECURSION = YES;
+				CLANG_WARN_INT_CONVERSION = YES;
+				CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
+				CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
+				CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
+				CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
+				CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
+				CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
+				CLANG_WARN_STRICT_PROTOTYPES = YES;
+				CLANG_WARN_SUSPICIOUS_MOVE = YES;
+				CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
+				CLANG_WARN_UNREACHABLE_CODE = YES;
+				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
+				COPY_PHASE_STRIP = NO;
+				DEBUG_INFORMATION_FORMAT = dwarf;
+				ENABLE_STRICT_OBJC_MSGSEND = YES;
+				ENABLE_TESTABILITY = YES;
+				ENABLE_USER_SCRIPT_SANDBOXING = YES;
+				GCC_C_LANGUAGE_STANDARD = gnu17;
+				GCC_DYNAMIC_NO_PIC = NO;
+				GCC_NO_COMMON_BLOCKS = YES;
+				GCC_OPTIMIZATION_LEVEL = 0;
+				GCC_PREPROCESSOR_DEFINITIONS = (
+					"DEBUG=1",
+					"$(inherited)",
+				);
+				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
+				GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
+				GCC_WARN_UNDECLARED_SELECTOR = YES;
+				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
+				GCC_WARN_UNUSED_FUNCTION = YES;
+				GCC_WARN_UNUSED_VARIABLE = YES;
+				IPHONEOS_DEPLOYMENT_TARGET = 26.2;
+				LOCALIZATION_PREFERS_STRING_CATALOGS = YES;
+				MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE;
+				MTL_FAST_MATH = YES;
+				ONLY_ACTIVE_ARCH = YES;
+				SDKROOT = iphoneos;
+				SWIFT_ACTIVE_COMPILATION_CONDITIONS = "DEBUG $(inherited)";
+				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
+			};
+			name = Debug;
+		};
+		E09D168F2F44CA20009C51A3 /* Release */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ALWAYS_SEARCH_USER_PATHS = NO;
+				ASSETCATALOG_COMPILER_GENERATE_SWIFT_ASSET_SYMBOL_EXTENSIONS = YES;
+				CLANG_ANALYZER_NONNULL = YES;
+				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
+				CLANG_CXX_LANGUAGE_STANDARD = "gnu++20";
+				CLANG_ENABLE_MODULES = YES;
+				CLANG_ENABLE_OBJC_ARC = YES;
+				CLANG_ENABLE_OBJC_WEAK = YES;
+				CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
+				CLANG_WARN_BOOL_CONVERSION = YES;
+				CLANG_WARN_COMMA = YES;
+				CLANG_WARN_CONSTANT_CONVERSION = YES;
+				CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
+				CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
+				CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
+				CLANG_WARN_EMPTY_BODY = YES;
+				CLANG_WARN_ENUM_CONVERSION = YES;
+				CLANG_WARN_INFINITE_RECURSION = YES;
+				CLANG_WARN_INT_CONVERSION = YES;
+				CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
+				CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
+				CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
+				CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
+				CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
+				CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
+				CLANG_WARN_STRICT_PROTOTYPES = YES;
+				CLANG_WARN_SUSPICIOUS_MOVE = YES;
+				CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
+				CLANG_WARN_UNREACHABLE_CODE = YES;
+				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
+				COPY_PHASE_STRIP = NO;
+				DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym";
+				ENABLE_NS_ASSERTIONS = NO;
+				ENABLE_STRICT_OBJC_MSGSEND = YES;
+				ENABLE_USER_SCRIPT_SANDBOXING = YES;
+				GCC_C_LANGUAGE_STANDARD = gnu17;
+				GCC_NO_COMMON_BLOCKS = YES;
+				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
+				GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
+				GCC_WARN_UNDECLARED_SELECTOR = YES;
+				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
+				GCC_WARN_UNUSED_FUNCTION = YES;
+				GCC_WARN_UNUSED_VARIABLE = YES;
+				IPHONEOS_DEPLOYMENT_TARGET = 26.2;
+				LOCALIZATION_PREFERS_STRING_CATALOGS = YES;
+				MTL_ENABLE_DEBUG_INFO = NO;
+				MTL_FAST_MATH = YES;
+				SDKROOT = iphoneos;
+				SWIFT_COMPILATION_MODE = wholemodule;
+				VALIDATE_PRODUCT = YES;
+			};
+			name = Release;
+		};
+		E09D16912F44CA20009C51A3 /* Debug */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
+				CODE_SIGN_STYLE = Automatic;
+				CURRENT_PROJECT_VERSION = 1;
+				DEVELOPMENT_TEAM = 3M3M67U93M;
+				ENABLE_PREVIEWS = YES;
+				GENERATE_INFOPLIST_FILE = YES;
+				INFOPLIST_FILE = "EXO-iOS/Info.plist";
+				INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
+				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
+				INFOPLIST_KEY_UILaunchScreen_Generation = YES;
+				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+				LD_RUNPATH_SEARCH_PATHS = (
+					"$(inherited)",
+					"@executable_path/Frameworks",
+				);
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = "com.exo.EXO-iOS";
+				PRODUCT_NAME = "$(TARGET_NAME)";
+				STRING_CATALOG_GENERATE_SYMBOLS = YES;
+				SWIFT_APPROACHABLE_CONCURRENCY = YES;
+				SWIFT_DEFAULT_ACTOR_ISOLATION = MainActor;
+				SWIFT_EMIT_LOC_STRINGS = YES;
+				SWIFT_UPCOMING_FEATURE_MEMBER_IMPORT_VISIBILITY = YES;
+				SWIFT_VERSION = 5.0;
+				TARGETED_DEVICE_FAMILY = "1,2";
+			};
+			name = Debug;
+		};
+		E09D16922F44CA20009C51A3 /* Release */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
+				CODE_SIGN_STYLE = Automatic;
+				CURRENT_PROJECT_VERSION = 1;
+				DEVELOPMENT_TEAM = 3M3M67U93M;
+				ENABLE_PREVIEWS = YES;
+				GENERATE_INFOPLIST_FILE = YES;
+				INFOPLIST_FILE = "EXO-iOS/Info.plist";
+				INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
+				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
+				INFOPLIST_KEY_UILaunchScreen_Generation = YES;
+				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+				LD_RUNPATH_SEARCH_PATHS = (
+					"$(inherited)",
+					"@executable_path/Frameworks",
+				);
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = "com.exo.EXO-iOS";
+				PRODUCT_NAME = "$(TARGET_NAME)";
+				STRING_CATALOG_GENERATE_SYMBOLS = YES;
+				SWIFT_APPROACHABLE_CONCURRENCY = YES;
+				SWIFT_DEFAULT_ACTOR_ISOLATION = MainActor;
+				SWIFT_EMIT_LOC_STRINGS = YES;
+				SWIFT_UPCOMING_FEATURE_MEMBER_IMPORT_VISIBILITY = YES;
+				SWIFT_VERSION = 5.0;
+				TARGETED_DEVICE_FAMILY = "1,2";
+			};
+			name = Release;
+		};
+		E09D16942F44CA20009C51A3 /* Debug */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				BUNDLE_LOADER = "$(TEST_HOST)";
+				CODE_SIGN_STYLE = Automatic;
+				CURRENT_PROJECT_VERSION = 1;
+				GENERATE_INFOPLIST_FILE = YES;
+				IPHONEOS_DEPLOYMENT_TARGET = 26.2;
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = "com.exo.EXO-iOSTests";
+				PRODUCT_NAME = "$(TARGET_NAME)";
+				STRING_CATALOG_GENERATE_SYMBOLS = NO;
+				SWIFT_APPROACHABLE_CONCURRENCY = YES;
+				SWIFT_EMIT_LOC_STRINGS = NO;
+				SWIFT_UPCOMING_FEATURE_MEMBER_IMPORT_VISIBILITY = YES;
+				SWIFT_VERSION = 5.0;
+				TARGETED_DEVICE_FAMILY = "1,2";
+				TEST_HOST = "$(BUILT_PRODUCTS_DIR)/EXO-iOS.app/$(BUNDLE_EXECUTABLE_FOLDER_PATH)/EXO-iOS";
+			};
+			name = Debug;
+		};
+		E09D16952F44CA20009C51A3 /* Release */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				BUNDLE_LOADER = "$(TEST_HOST)";
+				CODE_SIGN_STYLE = Automatic;
+				CURRENT_PROJECT_VERSION = 1;
+				GENERATE_INFOPLIST_FILE = YES;
+				IPHONEOS_DEPLOYMENT_TARGET = 26.2;
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = "com.exo.EXO-iOSTests";
+				PRODUCT_NAME = "$(TARGET_NAME)";
+				STRING_CATALOG_GENERATE_SYMBOLS = NO;
+				SWIFT_APPROACHABLE_CONCURRENCY = YES;
+				SWIFT_EMIT_LOC_STRINGS = NO;
+				SWIFT_UPCOMING_FEATURE_MEMBER_IMPORT_VISIBILITY = YES;
+				SWIFT_VERSION = 5.0;
+				TARGETED_DEVICE_FAMILY = "1,2";
+				TEST_HOST = "$(BUILT_PRODUCTS_DIR)/EXO-iOS.app/$(BUNDLE_EXECUTABLE_FOLDER_PATH)/EXO-iOS";
+			};
+			name = Release;
+		};
+		E09D16972F44CA20009C51A3 /* Debug */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				CODE_SIGN_STYLE = Automatic;
+				CURRENT_PROJECT_VERSION = 1;
+				GENERATE_INFOPLIST_FILE = YES;
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = "com.exo.EXO-iOSUITests";
+				PRODUCT_NAME = "$(TARGET_NAME)";
+				STRING_CATALOG_GENERATE_SYMBOLS = NO;
+				SWIFT_APPROACHABLE_CONCURRENCY = YES;
+				SWIFT_EMIT_LOC_STRINGS = NO;
+				SWIFT_UPCOMING_FEATURE_MEMBER_IMPORT_VISIBILITY = YES;
+				SWIFT_VERSION = 5.0;
+				TARGETED_DEVICE_FAMILY = "1,2";
+				TEST_TARGET_NAME = "EXO-iOS";
+			};
+			name = Debug;
+		};
+		E09D16982F44CA20009C51A3 /* Release */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				CODE_SIGN_STYLE = Automatic;
+				CURRENT_PROJECT_VERSION = 1;
+				GENERATE_INFOPLIST_FILE = YES;
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = "com.exo.EXO-iOSUITests";
+				PRODUCT_NAME = "$(TARGET_NAME)";
+				STRING_CATALOG_GENERATE_SYMBOLS = NO;
+				SWIFT_APPROACHABLE_CONCURRENCY = YES;
+				SWIFT_EMIT_LOC_STRINGS = NO;
+				SWIFT_UPCOMING_FEATURE_MEMBER_IMPORT_VISIBILITY = YES;
+				SWIFT_VERSION = 5.0;
+				TARGETED_DEVICE_FAMILY = "1,2";
+				TEST_TARGET_NAME = "EXO-iOS";
+			};
+			name = Release;
+		};
+/* End XCBuildConfiguration section */
+
+/* Begin XCConfigurationList section */
+		E09D166A2F44CA1E009C51A3 /* Build configuration list for PBXProject "EXO-iOS" */ = {
+			isa = XCConfigurationList;
+			buildConfigurations = (
+				E09D168E2F44CA20009C51A3 /* Debug */,
+				E09D168F2F44CA20009C51A3 /* Release */,
+			);
+			defaultConfigurationIsVisible = 0;
+			defaultConfigurationName = Release;
+		};
+		E09D16902F44CA20009C51A3 /* Build configuration list for PBXNativeTarget "EXO-iOS" */ = {
+			isa = XCConfigurationList;
+			buildConfigurations = (
+				E09D16912F44CA20009C51A3 /* Debug */,
+				E09D16922F44CA20009C51A3 /* Release */,
+			);
+			defaultConfigurationIsVisible = 0;
+			defaultConfigurationName = Release;
+		};
+		E09D16932F44CA20009C51A3 /* Build configuration list for PBXNativeTarget "EXO-iOSTests" */ = {
+			isa = XCConfigurationList;
+			buildConfigurations = (
+				E09D16942F44CA20009C51A3 /* Debug */,
+				E09D16952F44CA20009C51A3 /* Release */,
+			);
+			defaultConfigurationIsVisible = 0;
+			defaultConfigurationName = Release;
+		};
+		E09D16962F44CA20009C51A3 /* Build configuration list for PBXNativeTarget "EXO-iOSUITests" */ = {
+			isa = XCConfigurationList;
+			buildConfigurations = (
+				E09D16972F44CA20009C51A3 /* Debug */,
+				E09D16982F44CA20009C51A3 /* Release */,
+			);
+			defaultConfigurationIsVisible = 0;
+			defaultConfigurationName = Release;
+		};
+/* End XCConfigurationList section */
+
+/* Begin XCRemoteSwiftPackageReference section */
+		E09D17502F44F359009C51A3 /* XCRemoteSwiftPackageReference "mlx-swift-lm" */ = {
+			isa = XCRemoteSwiftPackageReference;
+			repositoryURL = "https://github.com/ml-explore/mlx-swift-lm";
+			requirement = {
+				kind = upToNextMajorVersion;
+				minimumVersion = 2.30.3;
+			};
+		};
+/* End XCRemoteSwiftPackageReference section */
+
+/* Begin XCSwiftPackageProductDependency section */
+		E09D17512F44F359009C51A3 /* MLXLLM */ = {
+			isa = XCSwiftPackageProductDependency;
+			package = E09D17502F44F359009C51A3 /* XCRemoteSwiftPackageReference "mlx-swift-lm" */;
+			productName = MLXLLM;
+		};
+		E09D17532F44F359009C51A3 /* MLXLMCommon */ = {
+			isa = XCSwiftPackageProductDependency;
+			package = E09D17502F44F359009C51A3 /* XCRemoteSwiftPackageReference "mlx-swift-lm" */;
+			productName = MLXLMCommon;
+		};
+/* End XCSwiftPackageProductDependency section */
+	};
+	rootObject = E09D16672F44CA1E009C51A3 /* Project object */;
+}
@@ -0,0 +1,7 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<Workspace
+   version = "1.0">
+   <FileRef
+      location = "self:">
+   </FileRef>
+</Workspace>
@@ -0,0 +1,60 @@
+{
+  "originHash" : "facc0ac7c70363ea20f6cd1235de91dea6b06f0d00190946045a6c8ae753abc2",
+  "pins" : [
+    {
+      "identity" : "mlx-swift",
+      "kind" : "remoteSourceControl",
+      "location" : "https://github.com/ml-explore/mlx-swift",
+      "state" : {
+        "revision" : "6ba4827fb82c97d012eec9ab4b2de21f85c3b33d",
+        "version" : "0.30.6"
+      }
+    },
+    {
+      "identity" : "mlx-swift-lm",
+      "kind" : "remoteSourceControl",
+      "location" : "https://github.com/ml-explore/mlx-swift-lm",
+      "state" : {
+        "revision" : "360c5052b81cc154b04ee0933597a4ad6db4b8ae",
+        "version" : "2.30.3"
+      }
+    },
+    {
+      "identity" : "swift-collections",
+      "kind" : "remoteSourceControl",
+      "location" : "https://github.com/apple/swift-collections.git",
+      "state" : {
+        "revision" : "7b847a3b7008b2dc2f47ca3110d8c782fb2e5c7e",
+        "version" : "1.3.0"
+      }
+    },
+    {
+      "identity" : "swift-jinja",
+      "kind" : "remoteSourceControl",
+      "location" : "https://github.com/huggingface/swift-jinja.git",
+      "state" : {
+        "revision" : "d81197f35f41445bc10e94600795e68c6f5e94b0",
+        "version" : "2.3.1"
+      }
+    },
+    {
+      "identity" : "swift-numerics",
+      "kind" : "remoteSourceControl",
+      "location" : "https://github.com/apple/swift-numerics",
+      "state" : {
+        "revision" : "0c0290ff6b24942dadb83a929ffaaa1481df04a2",
+        "version" : "1.1.1"
+      }
+    },
+    {
+      "identity" : "swift-transformers",
+      "kind" : "remoteSourceControl",
+      "location" : "https://github.com/huggingface/swift-transformers",
+      "state" : {
+        "revision" : "573e5c9036c2f136b3a8a071da8e8907322403d0",
+        "version" : "1.1.6"
+      }
+    }
+  ],
+  "version" : 3
+}
@@ -0,0 +1,20 @@
+{
+  "colors" : [
+    {
+      "color" : {
+        "color-space" : "srgb",
+        "components" : {
+          "alpha" : "1.000",
+          "blue" : "0x00",
+          "green" : "0xD7",
+          "red" : "0xFF"
+        }
+      },
+      "idiom" : "universal"
+    }
+  ],
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}
@@ -0,0 +1,38 @@
+{
+  "images" : [
+    {
+      "filename" : "AppIcon.png",
+      "idiom" : "universal",
+      "platform" : "ios",
+      "size" : "1024x1024"
+    },
+    {
+      "appearances" : [
+        {
+          "appearance" : "luminosity",
+          "value" : "dark"
+        }
+      ],
+      "filename" : "AppIcon.png",
+      "idiom" : "universal",
+      "platform" : "ios",
+      "size" : "1024x1024"
+    },
+    {
+      "appearances" : [
+        {
+          "appearance" : "luminosity",
+          "value" : "tinted"
+        }
+      ],
+      "filename" : "AppIcon.png",
+      "idiom" : "universal",
+      "platform" : "ios",
+      "size" : "1024x1024"
+    }
+  ],
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}
@@ -0,0 +1,6 @@
+{
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}
@@ -0,0 +1,21 @@
+{
+  "images" : [
+    {
+      "filename" : "exo-logo.png",
+      "idiom" : "universal",
+      "scale" : "1x"
+    },
+    {
+      "idiom" : "universal",
+      "scale" : "2x"
+    },
+    {
+      "idiom" : "universal",
+      "scale" : "3x"
+    }
+  ],
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}
@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>com.apple.developer.kernel.increased-memory-limit</key>
+	<true/>
+</dict>
+</plist>
@@ -0,0 +1,67 @@
+import SwiftUI
+import UIKit
+
+@main
+struct EXO_iOSApp: App {
+    @State private var clusterService = ClusterService()
+    @State private var discoveryService = DiscoveryService()
+    @State private var localInferenceService = LocalInferenceService()
+    @State private var chatService: ChatService?
+
+    init() {
+        let darkGray = UIColor(red: 0x1F / 255.0, green: 0x1F / 255.0, blue: 0x1F / 255.0, alpha: 1)
+        let yellow = UIColor(red: 0xFF / 255.0, green: 0xD7 / 255.0, blue: 0x00 / 255.0, alpha: 1)
+
+        let navAppearance = UINavigationBarAppearance()
+        navAppearance.configureWithOpaqueBackground()
+        navAppearance.backgroundColor = darkGray
+        navAppearance.titleTextAttributes = [
+            .foregroundColor: yellow,
+            .font: UIFont.monospacedSystemFont(ofSize: 17, weight: .semibold),
+        ]
+        navAppearance.largeTitleTextAttributes = [
+            .foregroundColor: yellow,
+            .font: UIFont.monospacedSystemFont(ofSize: 34, weight: .bold),
+        ]
+
+        UINavigationBar.appearance().standardAppearance = navAppearance
+        UINavigationBar.appearance().compactAppearance = navAppearance
+        UINavigationBar.appearance().scrollEdgeAppearance = navAppearance
+        UINavigationBar.appearance().tintColor = yellow
+    }
+
+    var body: some Scene {
+        WindowGroup {
+            if let chatService {
+                RootView()
+                    .environment(clusterService)
+                    .environment(discoveryService)
+                    .environment(chatService)
+                    .environment(localInferenceService)
+                    .preferredColorScheme(.dark)
+                    .task {
+                        await clusterService.attemptAutoReconnect()
+                        discoveryService.startBrowsing()
+                        await localInferenceService.prepareModel()
+                    }
+                    .onChange(of: discoveryService.discoveredClusters) { _, clusters in
+                        guard !clusterService.isConnected,
+                            case .disconnected = clusterService.connectionState,
+                            let first = clusters.first
+                        else { return }
+                        Task {
+                            await clusterService.connectToDiscoveredCluster(
+                                first, using: discoveryService)
+                        }
+                    }
+            } else {
+                Color.exoBlack.onAppear {
+                    chatService = ChatService(
+                        clusterService: clusterService,
+                        localInferenceService: localInferenceService
+                    )
+                }
+            }
+        }
+    }
+}
@@ -0,0 +1,19 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>UIUserInterfaceStyle</key>
+	<string>Dark</string>
+	<key>CFBundleDisplayName</key>
+	<string>EXO</string>
+	<key>NSLocalNetworkUsageDescription</key>
+	<string>EXO needs local network access to connect to your EXO cluster.</string>
+	<key>NSBonjourServices</key>
+	<array>
+		<string>_exo._tcp</string>
+		<string>_p2p._tcp</string>
+		<string>_p2p._udp</string>
+		<string>_libp2p._udp</string>
+	</array>
+</dict>
+</plist>
@@ -0,0 +1,129 @@
+import Foundation
+
+// MARK: - Request
+
+struct ChatCompletionRequest: Encodable {
+    let model: String
+    let messages: [ChatCompletionMessageParam]
+    let stream: Bool
+    let maxTokens: Int?
+    let temperature: Double?
+
+    enum CodingKeys: String, CodingKey {
+        case model, messages, stream, temperature
+        case maxTokens = "max_tokens"
+    }
+}
+
+struct ChatCompletionMessageParam: Encodable {
+    let role: String
+    let content: String
+}
+
+// MARK: - Streaming Response
+
+struct ChatCompletionChunk: Decodable {
+    let id: String
+    let model: String?
+    let choices: [StreamingChoice]
+    let usage: ChunkUsage?
+
+    init(id: String, model: String?, choices: [StreamingChoice], usage: ChunkUsage?) {
+        self.id = id
+        self.model = model
+        self.choices = choices
+        self.usage = usage
+    }
+}
+
+struct StreamingChoice: Decodable {
+    let index: Int
+    let delta: Delta
+    let finishReason: String?
+
+    enum CodingKeys: String, CodingKey {
+        case index, delta
+        case finishReason = "finish_reason"
+    }
+
+    init(index: Int, delta: Delta, finishReason: String?) {
+        self.index = index
+        self.delta = delta
+        self.finishReason = finishReason
+    }
+}
+
+struct Delta: Decodable {
+    let role: String?
+    let content: String?
+
+    init(role: String?, content: String?) {
+        self.role = role
+        self.content = content
+    }
+}
+
+struct ChunkUsage: Decodable {
+    let promptTokens: Int?
+    let completionTokens: Int?
+    let totalTokens: Int?
+
+    enum CodingKeys: String, CodingKey {
+        case promptTokens = "prompt_tokens"
+        case completionTokens = "completion_tokens"
+        case totalTokens = "total_tokens"
+    }
+
+    init(promptTokens: Int?, completionTokens: Int?, totalTokens: Int?) {
+        self.promptTokens = promptTokens
+        self.completionTokens = completionTokens
+        self.totalTokens = totalTokens
+    }
+}
+
+// MARK: - Non-Streaming Response
+
+struct ChatCompletionResponse: Decodable {
+    let id: String
+    let model: String?
+    let choices: [ResponseChoice]
+}
+
+struct ResponseChoice: Decodable {
+    let index: Int
+    let message: ResponseMessage
+    let finishReason: String?
+
+    enum CodingKeys: String, CodingKey {
+        case index, message
+        case finishReason = "finish_reason"
+    }
+}
+
+struct ResponseMessage: Decodable {
+    let role: String?
+    let content: String?
+}
+
+// MARK: - Models List
+
+struct ModelListResponse: Decodable {
+    let data: [ModelInfo]
+}
+
+struct ModelInfo: Decodable, Identifiable {
+    let id: String
+    let name: String?
+}
+
+// MARK: - Error
+
+struct APIErrorResponse: Decodable {
+    let error: APIErrorInfo
+}
+
+struct APIErrorInfo: Decodable {
+    let message: String
+    let type: String?
+    let code: Int?
+}
@@ -0,0 +1,26 @@
+import Foundation
+
+struct ChatMessage: Identifiable, Equatable {
+    let id: UUID
+    let role: Role
+    var content: String
+    let timestamp: Date
+    var isStreaming: Bool
+
+    enum Role: String, Codable {
+        case user
+        case assistant
+        case system
+    }
+
+    init(
+        id: UUID = UUID(), role: Role, content: String, timestamp: Date = Date(),
+        isStreaming: Bool = false
+    ) {
+        self.id = id
+        self.role = role
+        self.content = content
+        self.timestamp = timestamp
+        self.isStreaming = isStreaming
+    }
+}
@@ -0,0 +1,11 @@
+import Foundation
+
+struct ConnectionInfo: Codable, Equatable {
+    let host: String
+    let port: Int
+    let nodeId: String?
+
+    var baseURL: URL { URL(string: "http://\(host):\(port)")! }
+
+    static let defaultPort = 52415
+}
@@ -0,0 +1,34 @@
+import Foundation
+
+struct Conversation: Identifiable, Codable, Equatable {
+    let id: UUID
+    var title: String
+    var messages: [StoredMessage]
+    var modelId: String?
+    let createdAt: Date
+
+    init(
+        id: UUID = UUID(), title: String = "New Chat", messages: [StoredMessage] = [],
+        modelId: String? = nil, createdAt: Date = Date()
+    ) {
+        self.id = id
+        self.title = title
+        self.messages = messages
+        self.modelId = modelId
+        self.createdAt = createdAt
+    }
+}
+
+struct StoredMessage: Identifiable, Codable, Equatable {
+    let id: UUID
+    let role: String
+    var content: String
+    let timestamp: Date
+
+    init(id: UUID = UUID(), role: String, content: String, timestamp: Date = Date()) {
+        self.id = id
+        self.role = role
+        self.content = content
+        self.timestamp = timestamp
+    }
+}
@@ -0,0 +1,227 @@
+import Foundation
+
+@Observable
+@MainActor
+final class ChatService {
+    var conversations: [Conversation] = []
+    var activeConversationId: UUID?
+    private(set) var isGenerating: Bool = false
+    private var currentGenerationTask: Task<Void, Never>?
+
+    private let clusterService: ClusterService
+    private let localInferenceService: LocalInferenceService
+
+    var canSendMessage: Bool {
+        clusterService.isConnected || localInferenceService.isAvailable
+    }
+
+    var activeConversation: Conversation? {
+        guard let id = activeConversationId else { return nil }
+        return conversations.first { $0.id == id }
+    }
+
+    var activeMessages: [ChatMessage] {
+        guard let conversation = activeConversation else { return [] }
+        return conversation.messages.map { stored in
+            ChatMessage(
+                id: stored.id,
+                role: ChatMessage.Role(rawValue: stored.role) ?? .user,
+                content: stored.content,
+                timestamp: stored.timestamp
+            )
+        }
+    }
+
+    init(clusterService: ClusterService, localInferenceService: LocalInferenceService) {
+        self.clusterService = clusterService
+        self.localInferenceService = localInferenceService
+        loadConversations()
+    }
+
+    // MARK: - Conversation Management
+
+    func createConversation(modelId: String? = nil) {
+        let conversation = Conversation(
+            modelId: modelId ?? clusterService.availableModels.first?.id)
+        conversations.insert(conversation, at: 0)
+        activeConversationId = conversation.id
+        saveConversations()
+    }
+
+    func deleteConversation(id: UUID) {
+        conversations.removeAll { $0.id == id }
+        if activeConversationId == id {
+            activeConversationId = conversations.first?.id
+        }
+        saveConversations()
+    }
+
+    func setActiveConversation(id: UUID) {
+        activeConversationId = id
+    }
+
+    func setModelForActiveConversation(_ modelId: String) {
+        guard let index = conversations.firstIndex(where: { $0.id == activeConversationId }) else {
+            return
+        }
+        conversations[index].modelId = modelId
+        saveConversations()
+    }
+
+    // MARK: - Messaging
+
+    func sendMessage(_ text: String) {
+        guard !text.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty else { return }
+
+        if activeConversation == nil {
+            createConversation()
+        }
+
+        guard let index = conversations.firstIndex(where: { $0.id == activeConversationId }) else {
+            return
+        }
+
+        let userMessage = StoredMessage(role: "user", content: text)
+        conversations[index].messages.append(userMessage)
+
+        if conversations[index].title == "New Chat" {
+            let preview = String(text.prefix(40))
+            conversations[index].title = preview + (text.count > 40 ? "..." : "")
+        }
+
+        let modelId: String
+        if clusterService.isConnected {
+            guard
+                let clusterId = conversations[index].modelId
+                    ?? clusterService.availableModels.first?.id
+            else {
+                let errorMessage = StoredMessage(
+                    role: "assistant", content: "No model selected. Please select a model first.")
+                conversations[index].messages.append(errorMessage)
+                saveConversations()
+                return
+            }
+            modelId = clusterId
+        } else if localInferenceService.isAvailable {
+            modelId = localInferenceService.defaultModelId
+        } else {
+            let errorMessage = StoredMessage(
+                role: "assistant",
+                content: "Not connected to a cluster and local model is not available.")
+            conversations[index].messages.append(errorMessage)
+            saveConversations()
+            return
+        }
+
+        conversations[index].modelId = modelId
+
+        let assistantMessageId = UUID()
+        let assistantMessage = StoredMessage(
+            id: assistantMessageId, role: "assistant", content: "", timestamp: Date())
+        conversations[index].messages.append(assistantMessage)
+
+        let messagesForAPI = conversations[index].messages.dropLast().map { stored in
+            ChatCompletionMessageParam(role: stored.role, content: stored.content)
+        }
+
+        let request = ChatCompletionRequest(
+            model: modelId,
+            messages: Array(messagesForAPI),
+            stream: true,
+            maxTokens: 4096,
+            temperature: nil
+        )
+
+        let conversationId = conversations[index].id
+
+        isGenerating = true
+        currentGenerationTask = Task { [weak self] in
+            guard let self else { return }
+            await self.performStreaming(
+                request: request, conversationId: conversationId,
+                assistantMessageId: assistantMessageId)
+        }
+
+        saveConversations()
+    }
+
+    func cancelGeneration() {
+        currentGenerationTask?.cancel()
+        currentGenerationTask = nil
+        localInferenceService.cancelGeneration()
+        isGenerating = false
+    }
+
+    // MARK: - Streaming
+
+    private func performStreaming(
+        request: ChatCompletionRequest, conversationId: UUID, assistantMessageId: UUID
+    ) async {
+        defer {
+            isGenerating = false
+            currentGenerationTask = nil
+            saveConversations()
+        }
+
+        do {
+            let stream =
+                clusterService.isConnected
+                ? clusterService.streamChatCompletion(request: request)
+                : localInferenceService.streamChatCompletion(request: request)
+            for try await chunk in stream {
+                guard !Task.isCancelled else { return }
+                guard let content = chunk.choices.first?.delta.content, !content.isEmpty else {
+                    continue
+                }
+
+                if let convIndex = conversations.firstIndex(where: { $0.id == conversationId }),
+                    let msgIndex = conversations[convIndex].messages.firstIndex(where: {
+                        $0.id == assistantMessageId
+                    })
+                {
+                    conversations[convIndex].messages[msgIndex].content += content
+                }
+            }
+        } catch {
+            if !Task.isCancelled {
+                if let convIndex = conversations.firstIndex(where: { $0.id == conversationId }),
+                    let msgIndex = conversations[convIndex].messages.firstIndex(where: {
+                        $0.id == assistantMessageId
+                    })
+                {
+                    if conversations[convIndex].messages[msgIndex].content.isEmpty {
+                        conversations[convIndex].messages[msgIndex].content =
+                            "Error: \(error.localizedDescription)"
+                    }
+                }
+            }
+        }
+    }
+
+    // MARK: - Persistence
+
+    private static var storageURL: URL {
+        let documents = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)
+            .first!
+        return documents.appendingPathComponent("exo_conversations.json")
+    }
+
+    private func saveConversations() {
+        do {
+            let data = try JSONEncoder().encode(conversations)
+            try data.write(to: Self.storageURL, options: .atomic)
+        } catch {
+            // Save failed silently
+        }
+    }
+
+    private func loadConversations() {
+        do {
+            let data = try Data(contentsOf: Self.storageURL)
+            conversations = try JSONDecoder().decode([Conversation].self, from: data)
+            activeConversationId = conversations.first?.id
+        } catch {
+            conversations = []
+        }
+    }
+}
@@ -0,0 +1,246 @@
+import Foundation
+
+enum ConnectionState: Equatable {
+    case disconnected
+    case connecting
+    case connected(ConnectionInfo)
+}
+
+struct ModelOption: Identifiable, Equatable {
+    let id: String
+    let displayName: String
+}
+
+@Observable
+@MainActor
+final class ClusterService {
+    private(set) var connectionState: ConnectionState = .disconnected
+    private(set) var availableModels: [ModelOption] = []
+    private(set) var lastError: String?
+
+    private let session: URLSession
+    private let decoder: JSONDecoder
+    private var pollingTask: Task<Void, Never>?
+    private var heartbeatTask: Task<Void, Never>?
+
+    private static let connectionInfoKey = "exo_last_connection_info"
+
+    var isConnected: Bool {
+        if case .connected = connectionState { return true }
+        return false
+    }
+
+    var currentConnection: ConnectionInfo? {
+        if case .connected(let info) = connectionState { return info }
+        return nil
+    }
+
+    init(session: URLSession = .shared) {
+        self.session = session
+        let decoder = JSONDecoder()
+        self.decoder = decoder
+    }
+
+    // MARK: - Connection
+
+    func connect(to info: ConnectionInfo) async {
+        connectionState = .connecting
+        lastError = nil
+
+        do {
+            let url = info.baseURL.appendingPathComponent("node_id")
+            var request = URLRequest(url: url)
+            request.timeoutInterval = 5
+            request.cachePolicy = .reloadIgnoringLocalCacheData
+            let (_, response) = try await session.data(for: request)
+
+            guard let httpResponse = response as? HTTPURLResponse,
+                (200..<300).contains(httpResponse.statusCode)
+            else {
+                throw URLError(.badServerResponse)
+            }
+
+            connectionState = .connected(info)
+            persistConnection(info)
+            startPolling()
+            startHeartbeat()
+            await fetchModels(baseURL: info.baseURL)
+        } catch {
+            connectionState = .disconnected
+            lastError = "Could not connect to \(info.host):\(info.port)"
+        }
+    }
+
+    func connectToDiscoveredCluster(
+        _ cluster: DiscoveredCluster, using discoveryService: DiscoveryService
+    ) async {
+        guard case .disconnected = connectionState else { return }
+        connectionState = .connecting
+        lastError = nil
+
+        guard let info = await discoveryService.resolve(cluster) else {
+            connectionState = .disconnected
+            lastError = "Could not resolve \(cluster.name)"
+            return
+        }
+        connectionState = .disconnected  // reset so connect() can proceed
+        await connect(to: info)
+    }
+
+    func disconnect() {
+        stopPolling()
+        stopHeartbeat()
+        connectionState = .disconnected
+        availableModels = []
+        lastError = nil
+    }
+
+    func attemptAutoReconnect() async {
+        guard case .disconnected = connectionState,
+            let info = loadPersistedConnection()
+        else { return }
+        await connect(to: info)
+    }
+
+    // MARK: - Polling
+
+    private func startPolling(interval: TimeInterval = 2.0) {
+        stopPolling()
+        pollingTask = Task { [weak self] in
+            while !Task.isCancelled {
+                try? await Task.sleep(for: .seconds(interval))
+                guard let self, !Task.isCancelled else { return }
+                guard let connection = self.currentConnection else { return }
+                await self.fetchModels(baseURL: connection.baseURL)
+            }
+        }
+    }
+
+    private func stopPolling() {
+        pollingTask?.cancel()
+        pollingTask = nil
+    }
+
+    // MARK: - Heartbeat
+
+    private func startHeartbeat(interval: TimeInterval = 10.0) {
+        stopHeartbeat()
+        heartbeatTask = Task { [weak self] in
+            // Send immediately, then on interval
+            if let self, let connection = self.currentConnection {
+                await self.sendHeartbeat(baseURL: connection.baseURL)
+            }
+            while !Task.isCancelled {
+                try? await Task.sleep(for: .seconds(interval))
+                guard let self, !Task.isCancelled else { return }
+                guard let connection = self.currentConnection else { return }
+                await self.sendHeartbeat(baseURL: connection.baseURL)
+            }
+        }
+    }
+
+    private func stopHeartbeat() {
+        heartbeatTask?.cancel()
+        heartbeatTask = nil
+    }
+
+    private func sendHeartbeat(baseURL: URL) async {
+        do {
+            let deviceInfo = DeviceInfoService.gather()
+            let url = baseURL.appendingPathComponent("v1/lite_node/heartbeat")
+            var request = URLRequest(url: url)
+            request.httpMethod = "POST"
+            request.setValue("application/json", forHTTPHeaderField: "Content-Type")
+            request.timeoutInterval = 5
+            request.httpBody = try JSONEncoder().encode(deviceInfo)
+            let (_, response) = try await session.data(for: request)
+
+            if let httpResponse = response as? HTTPURLResponse,
+               !(200..<300).contains(httpResponse.statusCode) {
+                // Heartbeat failed silently — will retry on next interval
+            }
+        } catch {
+            // Heartbeat failed silently — will retry on next interval
+        }
+    }
+
+    // MARK: - API
+
+    private func fetchModels(baseURL: URL) async {
+        do {
+            let url = baseURL.appendingPathComponent("models")
+            var request = URLRequest(url: url)
+            request.cachePolicy = .reloadIgnoringLocalCacheData
+            let (data, response) = try await session.data(for: request)
+
+            guard let httpResponse = response as? HTTPURLResponse,
+                (200..<300).contains(httpResponse.statusCode)
+            else { return }
+
+            let list = try decoder.decode(ModelListResponse.self, from: data)
+            availableModels = list.data.map {
+                ModelOption(id: $0.id, displayName: $0.name ?? $0.id)
+            }
+        } catch {
+            // Models fetch failed silently — will retry on next poll
+        }
+    }
+
+    func streamChatCompletion(request body: ChatCompletionRequest) -> AsyncThrowingStream<
+        ChatCompletionChunk, Error
+    > {
+        AsyncThrowingStream { continuation in
+            let task = Task { [weak self] in
+                guard let self, let connection = self.currentConnection else {
+                    continuation.finish(throwing: URLError(.notConnectedToInternet))
+                    return
+                }
+
+                do {
+                    let url = connection.baseURL.appendingPathComponent("v1/chat/completions")
+                    var request = URLRequest(url: url)
+                    request.httpMethod = "POST"
+                    request.setValue("application/json", forHTTPHeaderField: "Content-Type")
+                    request.httpBody = try JSONEncoder().encode(body)
+
+                    let (bytes, response) = try await self.session.bytes(for: request)
+
+                    guard let httpResponse = response as? HTTPURLResponse,
+                        (200..<300).contains(httpResponse.statusCode)
+                    else {
+                        continuation.finish(throwing: URLError(.badServerResponse))
+                        return
+                    }
+
+                    let parser = SSEStreamParser<ChatCompletionChunk>(
+                        bytes: bytes, decoder: self.decoder)
+                    for try await chunk in parser {
+                        continuation.yield(chunk)
+                    }
+                    continuation.finish()
+                } catch {
+                    continuation.finish(throwing: error)
+                }
+            }
+
+            continuation.onTermination = { _ in
+                task.cancel()
+            }
+        }
+    }
+
+    // MARK: - Persistence
+
+    private func persistConnection(_ info: ConnectionInfo) {
+        if let data = try? JSONEncoder().encode(info) {
+            UserDefaults.standard.set(data, forKey: Self.connectionInfoKey)
+        }
+    }
+
+    private func loadPersistedConnection() -> ConnectionInfo? {
+        guard let data = UserDefaults.standard.data(forKey: Self.connectionInfoKey) else {
+            return nil
+        }
+        return try? JSONDecoder().decode(ConnectionInfo.self, from: data)
+    }
+}
@@ -0,0 +1,115 @@
+import Foundation
+import os
+import UIKit
+
+struct DeviceInfo: Encodable {
+    let nodeId: String
+    let model: String
+    let chip: String
+    let osVersion: String
+    let friendlyName: String
+    let ramTotal: Int
+    let ramAvailable: Int
+
+    enum CodingKeys: String, CodingKey {
+        case nodeId = "node_id"
+        case model
+        case chip
+        case osVersion = "os_version"
+        case friendlyName = "friendly_name"
+        case ramTotal = "ram_total"
+        case ramAvailable = "ram_available"
+    }
+}
+
+enum DeviceInfoService {
+    private static let liteNodeIdKey = "exo_lite_node_id"
+
+    static var liteNodeId: String {
+        if let existing = UserDefaults.standard.string(forKey: liteNodeIdKey) {
+            return existing
+        }
+        let newId = UUID().uuidString.lowercased()
+        UserDefaults.standard.set(newId, forKey: liteNodeIdKey)
+        return newId
+    }
+
+    static func gather() -> DeviceInfo {
+        let model = modelName()
+        let chip = chipName(for: model)
+
+        let totalRam = Int(ProcessInfo.processInfo.physicalMemory)
+        let availableRam = availableMemory()
+
+        return DeviceInfo(
+            nodeId: liteNodeId,
+            model: model,
+            chip: chip,
+            osVersion: UIDevice.current.systemVersion,
+            friendlyName: UIDevice.current.name,
+            ramTotal: totalRam,
+            ramAvailable: availableRam
+        )
+    }
+
+    // MARK: - Private
+
+    private static func modelName() -> String {
+        var systemInfo = utsname()
+        uname(&systemInfo)
+        let machine = withUnsafePointer(to: &systemInfo.machine) {
+            $0.withMemoryRebound(to: CChar.self, capacity: 1) {
+                String(cString: $0)
+            }
+        }
+        return modelMapping[machine] ?? machine
+    }
+
+    private static func chipName(for model: String) -> String {
+        let lower = model.lowercased()
+        if lower.contains("iphone 16 pro") || lower.contains("iphone 16 pro max") {
+            return "Apple A18 Pro"
+        } else if lower.contains("iphone 16") {
+            return "Apple A18"
+        } else if lower.contains("iphone 15 pro") || lower.contains("iphone 15 pro max") {
+            return "Apple A17 Pro"
+        } else if lower.contains("iphone 15") {
+            return "Apple A16 Bionic"
+        } else if lower.contains("iphone 14 pro") || lower.contains("iphone 14 pro max") {
+            return "Apple A16 Bionic"
+        } else if lower.contains("iphone 14") {
+            return "Apple A15 Bionic"
+        }
+        return "Apple Silicon"
+    }
+
+    private static func availableMemory() -> Int {
+        return Int(os_proc_available_memory())
+    }
+
+    private static let modelMapping: [String: String] = [
+        // iPhone 16 series
+        "iPhone17,1": "iPhone 16 Pro",
+        "iPhone17,2": "iPhone 16 Pro Max",
+        "iPhone17,3": "iPhone 16",
+        "iPhone17,4": "iPhone 16 Plus",
+        // iPhone 15 series
+        "iPhone16,1": "iPhone 15 Pro",
+        "iPhone16,2": "iPhone 15 Pro Max",
+        "iPhone15,4": "iPhone 15",
+        "iPhone15,5": "iPhone 15 Plus",
+        // iPhone 14 series
+        "iPhone15,2": "iPhone 14 Pro",
+        "iPhone15,3": "iPhone 14 Pro Max",
+        "iPhone14,7": "iPhone 14",
+        "iPhone14,8": "iPhone 14 Plus",
+        // iPhone 13 series
+        "iPhone14,2": "iPhone 13 Pro",
+        "iPhone14,3": "iPhone 13 Pro Max",
+        "iPhone14,5": "iPhone 13",
+        "iPhone14,4": "iPhone 13 mini",
+        // Simulator
+        "arm64": "iPhone (Simulator)",
+        "x86_64": "iPhone (Simulator)",
+    ]
+}
@@ -0,0 +1,123 @@
+import Foundation
+import Network
+import os
+
+struct DiscoveredCluster: Identifiable, Equatable {
+    let id: String
+    let name: String
+    let endpoint: NWEndpoint
+
+    static func == (lhs: DiscoveredCluster, rhs: DiscoveredCluster) -> Bool {
+        lhs.id == rhs.id && lhs.name == rhs.name
+    }
+}
+
+@Observable
+@MainActor
+final class DiscoveryService {
+    private(set) var discoveredClusters: [DiscoveredCluster] = []
+    private(set) var isSearching = false
+
+    private var browser: NWBrowser?
+
+    func startBrowsing() {
+        guard browser == nil else { return }
+
+        let browser = NWBrowser(for: .bonjour(type: "_exo._tcp", domain: nil), using: .tcp)
+
+        browser.stateUpdateHandler = { [weak self] state in
+            guard let service = self else { return }
+            Task { @MainActor in
+                switch state {
+                case .ready:
+                    service.isSearching = true
+                case .failed, .cancelled:
+                    service.isSearching = false
+                default:
+                    break
+                }
+            }
+        }
+
+        browser.browseResultsChangedHandler = { [weak self] results, _ in
+            guard let service = self else { return }
+            Task { @MainActor in
+                service.discoveredClusters = results.compactMap { result in
+                    guard case .service(let name, _, _, _) = result.endpoint else {
+                        return nil
+                    }
+                    return DiscoveredCluster(
+                        id: name,
+                        name: name,
+                        endpoint: result.endpoint
+                    )
+                }
+            }
+        }
+
+        browser.start(queue: .main)
+        self.browser = browser
+    }
+
+    func stopBrowsing() {
+        browser?.cancel()
+        browser = nil
+        isSearching = false
+        discoveredClusters = []
+    }
+
+    /// Resolve a discovered Bonjour endpoint to an IP address and port, then return a ConnectionInfo.
+    func resolve(_ cluster: DiscoveredCluster) async -> ConnectionInfo? {
+        await withCheckedContinuation { continuation in
+            let didResume = OSAllocatedUnfairLock(initialState: false)
+            let connection = NWConnection(to: cluster.endpoint, using: .tcp)
+            connection.stateUpdateHandler = { state in
+                guard
+                    didResume.withLock({
+                        guard !$0 else { return false }
+                        $0 = true
+                        return true
+                    })
+                else { return }
+                switch state {
+                case .ready:
+                    if let innerEndpoint = connection.currentPath?.remoteEndpoint,
+                        case .hostPort(let host, let port) = innerEndpoint
+                    {
+                        var hostString: String
+                        switch host {
+                        case .ipv4(let addr):
+                            hostString = "\(addr)"
+                        case .ipv6(let addr):
+                            hostString = "\(addr)"
+                        case .name(let name, _):
+                            hostString = name
+                        @unknown default:
+                            hostString = "\(host)"
+                        }
+                        // Strip interface scope suffix (e.g. "%en0")
+                        if let pct = hostString.firstIndex(of: "%") {
+                            hostString = String(hostString[..<pct])
+                        }
+                        let info = ConnectionInfo(
+                            host: hostString,
+                            port: Int(port.rawValue),
+                            nodeId: nil
+                        )
+                        connection.cancel()
+                        continuation.resume(returning: info)
+                    } else {
+                        connection.cancel()
+                        continuation.resume(returning: nil)
+                    }
+                case .failed, .cancelled:
+                    continuation.resume(returning: nil)
+                default:
+                    // Not a terminal state — allow future callbacks
+                    didResume.withLock { $0 = false }
+                }
+            }
+            connection.start(queue: .global(qos: .userInitiated))
+        }
+    }
+}
@@ -0,0 +1,201 @@
+import Foundation
+import MLXLLM
+import MLXLMCommon
+
+enum LocalModelState: Equatable {
+    case notDownloaded
+    case downloading(progress: Double)
+    case downloaded
+    case loading
+    case ready
+    case generating
+    case error(String)
+}
+
+@Observable
+@MainActor
+final class LocalInferenceService {
+    private(set) var modelState: LocalModelState = .notDownloaded
+    private var modelContainer: ModelContainer?
+    private var generationTask: Task<Void, Never>?
+
+    let defaultModelId = "mlx-community/Qwen3-0.6B-4bit"
+
+    private static let modelDownloadedKey = "exo_local_model_downloaded"
+
+    var isReady: Bool {
+        modelState == .ready
+    }
+
+    var isAvailable: Bool {
+        modelState == .ready || modelState == .generating
+    }
+
+    init() {
+        if UserDefaults.standard.bool(forKey: Self.modelDownloadedKey) {
+            modelState = .downloaded
+        }
+    }
+
+    // MARK: - Model Lifecycle
+
+    func prepareModel() async {
+        guard modelState == .notDownloaded || modelState == .downloaded else { return }
+
+        let wasDownloaded = modelState == .downloaded
+
+        if !wasDownloaded {
+            modelState = .downloading(progress: 0)
+        } else {
+            modelState = .loading
+        }
+
+        do {
+            let container = try await loadModelContainer(
+                id: defaultModelId
+            ) { [weak self] progress in
+                guard let self else { return }
+                Task { @MainActor in
+                    if case .downloading = self.modelState {
+                        self.modelState = .downloading(progress: progress.fractionCompleted)
+                    }
+                }
+            }
+
+            self.modelContainer = container
+            UserDefaults.standard.set(true, forKey: Self.modelDownloadedKey)
+            modelState = .ready
+        } catch {
+            modelState = .error(error.localizedDescription)
+        }
+    }
+
+    func unloadModel() {
+        cancelGeneration()
+        modelContainer = nil
+        modelState = .downloaded
+    }
+
+    // MARK: - Generation
+
+    func streamChatCompletion(request: ChatCompletionRequest) -> AsyncThrowingStream<
+        ChatCompletionChunk, Error
+    > {
+        AsyncThrowingStream { continuation in
+            let task = Task { [weak self] in
+                guard let self else {
+                    continuation.finish(throwing: LocalInferenceError.serviceUnavailable)
+                    return
+                }
+
+                guard let container = self.modelContainer else {
+                    continuation.finish(throwing: LocalInferenceError.modelNotLoaded)
+                    return
+                }
+
+                await MainActor.run {
+                    self.modelState = .generating
+                }
+
+                defer {
+                    Task { @MainActor [weak self] in
+                        if self?.modelState == .generating {
+                            self?.modelState = .ready
+                        }
+                    }
+                }
+
+                let chunkId = "local-\(UUID().uuidString)"
+
+                do {
+                    // Build Chat.Message array from the request
+                    var chatMessages: [Chat.Message] = []
+                    for msg in request.messages {
+                        switch msg.role {
+                        case "system":
+                            chatMessages.append(.system(msg.content))
+                        case "assistant":
+                            chatMessages.append(.assistant(msg.content))
+                        default:
+                            chatMessages.append(.user(msg.content))
+                        }
+                    }
+
+                    // Use ChatSession for streaming generation
+                    let session = ChatSession(
+                        container,
+                        history: chatMessages,
+                        generateParameters: GenerateParameters(
+                            maxTokens: request.maxTokens ?? 4096,
+                            temperature: Float(request.temperature ?? 0.7)
+                        )
+                    )
+
+                    // Stream with an empty prompt since history already contains the conversation
+                    let stream = session.streamResponse(to: "")
+                    for try await text in stream {
+                        if Task.isCancelled { break }
+
+                        let chunk = ChatCompletionChunk(
+                            id: chunkId,
+                            model: request.model,
+                            choices: [
+                                StreamingChoice(
+                                    index: 0,
+                                    delta: Delta(role: nil, content: text),
+                                    finishReason: nil
+                                )
+                            ],
+                            usage: nil
+                        )
+                        continuation.yield(chunk)
+                    }
+
+                    // Send final chunk with finish reason
+                    let finalChunk = ChatCompletionChunk(
+                        id: chunkId,
+                        model: request.model,
+                        choices: [
+                            StreamingChoice(
+                                index: 0,
+                                delta: Delta(role: nil, content: nil),
+                                finishReason: "stop"
+                            )
+                        ],
+                        usage: nil
+                    )
+                    continuation.yield(finalChunk)
+                    continuation.finish()
+                } catch {
+                    continuation.finish(throwing: error)
+                }
+            }
+
+            self.generationTask = task
+
+            continuation.onTermination = { _ in
+                task.cancel()
+            }
+        }
+    }
+
+    func cancelGeneration() {
+        generationTask?.cancel()
+        generationTask = nil
+        if modelState == .generating {
+            modelState = .ready
+        }
+    }
+}
+
+enum LocalInferenceError: LocalizedError {
+    case serviceUnavailable
+    case modelNotLoaded
+
+    var errorDescription: String? {
+        switch self {
+        case .serviceUnavailable: "Local inference service is unavailable"
+        case .modelNotLoaded: "Local model is not loaded"
+        }
+    }
+}
@@ -0,0 +1,50 @@
+import Foundation
+
+struct SSEStreamParser<T: Decodable>: AsyncSequence {
+    typealias Element = T
+
+    let bytes: URLSession.AsyncBytes
+    let decoder: JSONDecoder
+
+    init(bytes: URLSession.AsyncBytes, decoder: JSONDecoder = JSONDecoder()) {
+        self.bytes = bytes
+        self.decoder = decoder
+    }
+
+    func makeAsyncIterator() -> AsyncIterator {
+        AsyncIterator(lines: bytes.lines, decoder: decoder)
+    }
+
+    struct AsyncIterator: AsyncIteratorProtocol {
+        var lines: AsyncLineSequence<URLSession.AsyncBytes>.AsyncIterator
+        let decoder: JSONDecoder
+
+        init(lines: AsyncLineSequence<URLSession.AsyncBytes>, decoder: JSONDecoder) {
+            self.lines = lines.makeAsyncIterator()
+            self.decoder = decoder
+        }
+
+        mutating func next() async throws -> T? {
+            while let line = try await lines.next() {
+                let trimmed = line.trimmingCharacters(in: .whitespaces)
+
+                guard trimmed.hasPrefix("data: ") else { continue }
+
+                let payload = String(trimmed.dropFirst(6))
+
+                if payload == "[DONE]" {
+                    return nil
+                }
+
+                guard let data = payload.data(using: .utf8) else { continue }
+
+                do {
+                    return try decoder.decode(T.self, from: data)
+                } catch {
+                    continue
+                }
+            }
+            return nil
+        }
+    }
+}
@@ -0,0 +1,203 @@
+import SwiftUI
+
+struct ChatView: View {
+    @Environment(ClusterService.self) private var clusterService
+    @Environment(ChatService.self) private var chatService
+    @Environment(LocalInferenceService.self) private var localInferenceService
+    @State private var inputText = ""
+    @State private var showModelSelector = false
+
+    var body: some View {
+        VStack(spacing: 0) {
+            modelBar
+
+            GradientDivider()
+
+            messageList
+
+            GradientDivider()
+
+            inputBar
+        }
+        .background(Color.exoBlack)
+        .sheet(isPresented: $showModelSelector) {
+            ModelSelectorView(
+                models: clusterService.availableModels,
+                selectedModelId: chatService.activeConversation?.modelId
+            ) { modelId in
+                chatService.setModelForActiveConversation(modelId)
+            }
+            .presentationBackground(Color.exoDarkGray)
+        }
+    }
+
+    // MARK: - Model Bar
+
+    private var useLocalModel: Bool {
+        !clusterService.isConnected && localInferenceService.isAvailable
+    }
+
+    private var modelBar: some View {
+        Button {
+            if !useLocalModel {
+                showModelSelector = true
+            }
+        } label: {
+            HStack {
+                Image(systemName: useLocalModel ? "iphone" : "cpu")
+                    .font(.exoCaption)
+                    .foregroundStyle(useLocalModel ? Color.exoYellow : Color.exoLightGray)
+
+                if useLocalModel {
+                    Text(localInferenceService.defaultModelId)
+                        .font(.exoSubheadline)
+                        .foregroundStyle(Color.exoForeground)
+                        .lineLimit(1)
+                } else if let modelId = chatService.activeConversation?.modelId {
+                    Text(modelId)
+                        .font(.exoSubheadline)
+                        .foregroundStyle(Color.exoForeground)
+                        .lineLimit(1)
+                } else {
+                    Text("SELECT MODEL")
+                        .font(.exoSubheadline)
+                        .tracking(1.5)
+                        .foregroundStyle(Color.exoLightGray)
+                }
+
+                Spacer()
+
+                if useLocalModel {
+                    Text("ON-DEVICE")
+                        .font(.exoCaption)
+                        .tracking(1)
+                        .foregroundStyle(Color.exoYellow)
+                        .padding(.horizontal, 6)
+                        .padding(.vertical, 2)
+                        .background(Color.exoYellow.opacity(0.15))
+                        .clipShape(Capsule())
+                } else {
+                    Image(systemName: "chevron.right")
+                        .font(.caption)
+                        .foregroundStyle(Color.exoLightGray)
+                }
+            }
+            .padding(.horizontal)
+            .padding(.vertical, 10)
+            .background(Color.exoDarkGray)
+        }
+        .tint(.primary)
+        .disabled(useLocalModel)
+    }
+
+    // MARK: - Messages
+
+    private var messageList: some View {
+        ScrollViewReader { proxy in
+            ScrollView {
+                LazyVStack(spacing: 12) {
+                    if chatService.activeMessages.isEmpty {
+                        emptyState
+                    } else {
+                        ForEach(chatService.activeMessages) { message in
+                            MessageBubbleView(message: message)
+                                .id(message.id)
+                        }
+                    }
+                }
+                .padding()
+            }
+            .background(Color.exoBlack)
+            .onChange(of: chatService.activeMessages.last?.content) {
+                if let lastId = chatService.activeMessages.last?.id {
+                    withAnimation(.easeOut(duration: 0.2)) {
+                        proxy.scrollTo(lastId, anchor: .bottom)
+                    }
+                }
+            }
+        }
+    }
+
+    private var emptyState: some View {
+        VStack(spacing: 16) {
+            Spacer(minLength: 80)
+
+            ZStack {
+                Circle()
+                    .stroke(Color.exoYellow.opacity(0.15), lineWidth: 1)
+                    .frame(width: 80, height: 80)
+                Circle()
+                    .stroke(Color.exoYellow.opacity(0.3), lineWidth: 1)
+                    .frame(width: 56, height: 56)
+                Circle()
+                    .fill(Color.exoYellow.opacity(0.15))
+                    .frame(width: 32, height: 32)
+                Circle()
+                    .fill(Color.exoYellow)
+                    .frame(width: 8, height: 8)
+                    .shadow(color: Color.exoYellow.opacity(0.6), radius: 6)
+            }
+
+            Text("AWAITING INPUT")
+                .font(.exoSubheadline)
+                .tracking(3)
+                .foregroundStyle(Color.exoLightGray)
+
+            Text("Send a message to begin.")
+                .font(.exoCaption)
+                .foregroundStyle(Color.exoLightGray.opacity(0.6))
+
+            Spacer(minLength: 80)
+        }
+        .padding()
+    }
+
+    // MARK: - Input
+
+    private var inputBar: some View {
+        HStack(alignment: .bottom, spacing: 8) {
+            TextField("Message...", text: $inputText, axis: .vertical)
+                .font(.exoBody)
+                .lineLimit(1...6)
+                .textFieldStyle(.plain)
+                .padding(10)
+                .background(Color.exoMediumGray)
+                .foregroundStyle(Color.exoForeground)
+                .clipShape(RoundedRectangle(cornerRadius: 8))
+
+            if chatService.isGenerating {
+                Button {
+                    chatService.cancelGeneration()
+                } label: {
+                    Image(systemName: "stop.circle.fill")
+                        .font(.title2)
+                        .foregroundStyle(Color.exoDestructive)
+                }
+            } else {
+                Button {
+                    let text = inputText
+                    inputText = ""
+                    chatService.sendMessage(text)
+                } label: {
+                    Text("SEND")
+                        .font(.exoMono(12, weight: .bold))
+                        .tracking(1)
+                        .foregroundStyle(canSend ? Color.exoBlack : Color.exoLightGray)
+                        .padding(.horizontal, 14)
+                        .padding(.vertical, 8)
+                        .background(canSend ? Color.exoYellow : Color.exoMediumGray)
+                        .clipShape(RoundedRectangle(cornerRadius: 8))
+                }
+                .disabled(!canSend)
+            }
+        }
+        .padding(.horizontal)
+        .padding(.vertical, 8)
+        .background(Color.exoDarkGray)
+    }
+
+    private var canSend: Bool {
+        !inputText.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
+            && (clusterService.isConnected || localInferenceService.isAvailable)
+    }
+}
@@ -0,0 +1,54 @@
+import SwiftUI
+
+struct MessageBubbleView: View {
+    let message: ChatMessage
+
+    private var isAssistant: Bool { message.role == .assistant }
+
+    var body: some View {
+        HStack {
+            if message.role == .user { Spacer(minLength: 48) }
+
+            VStack(alignment: isAssistant ? .leading : .trailing, spacing: 6) {
+                // Header
+                HStack(spacing: 4) {
+                    if isAssistant {
+                        Circle()
+                            .fill(Color.exoYellow)
+                            .frame(width: 6, height: 6)
+                            .shadow(color: Color.exoYellow.opacity(0.6), radius: 4)
+                        Text("EXO")
+                            .font(.exoMono(10, weight: .bold))
+                            .tracking(1.5)
+                            .foregroundStyle(Color.exoYellow)
+                    } else {
+                        Text("QUERY")
+                            .font(.exoMono(10, weight: .medium))
+                            .tracking(1.5)
+                            .foregroundStyle(Color.exoLightGray)
+                    }
+                }
+
+                // Bubble
+                HStack(spacing: 0) {
+                    if isAssistant {
+                        RoundedRectangle(cornerRadius: 1)
+                            .fill(Color.exoYellow.opacity(0.5))
+                            .frame(width: 2)
+                    }
+
+                    Text(message.content + (message.isStreaming ? " \u{258C}" : ""))
+                        .font(.exoBody)
+                        .textSelection(.enabled)
+                        .foregroundStyle(Color.exoForeground)
+                        .padding(.horizontal, 14)
+                        .padding(.vertical, 10)
+                }
+                .background(Color.exoDarkGray)
+                .clipShape(RoundedRectangle(cornerRadius: 8))
+            }
+
+            if isAssistant { Spacer(minLength: 48) }
+        }
+    }
+}
@@ -0,0 +1,75 @@
+import SwiftUI
+
+struct ModelSelectorView: View {
+    let models: [ModelOption]
+    let selectedModelId: String?
+    let onSelect: (String) -> Void
+    @Environment(\.dismiss) private var dismiss
+
+    var body: some View {
+        NavigationStack {
+            List {
+                if models.isEmpty {
+                    emptyContent
+                } else {
+                    modelsList
+                }
+            }
+            .scrollContentBackground(.hidden)
+            .background(Color.exoBlack)
+            .navigationTitle("SELECT MODEL")
+            .navigationBarTitleDisplayMode(.inline)
+            .toolbar {
+                ToolbarItem(placement: .cancellationAction) {
+                    Button("Cancel") { dismiss() }
+                        .font(.exoSubheadline)
+                        .foregroundStyle(Color.exoYellow)
+                }
+            }
+        }
+    }
+
+    private var emptyContent: some View {
+        ContentUnavailableView(
+            "No Models Available",
+            systemImage: "cpu",
+            description: Text("Connect to an EXO cluster to see available models.")
+        )
+        .foregroundStyle(Color.exoLightGray)
+        .listRowBackground(Color.exoBlack)
+    }
+
+    private var modelsList: some View {
+        ForEach(models) { model in
+            Button {
+                onSelect(model.id)
+                dismiss()
+            } label: {
+                modelRow(model)
+            }
+            .tint(.primary)
+            .listRowBackground(Color.exoDarkGray)
+        }
+    }
+
+    private func modelRow(_ model: ModelOption) -> some View {
+        HStack {
+            VStack(alignment: .leading, spacing: 2) {
+                Text(model.displayName)
+                    .font(.exoSubheadline)
+                    .foregroundStyle(Color.exoForeground)
+                Text(model.id)
+                    .font(.exoCaption)
+                    .foregroundStyle(Color.exoLightGray)
+            }
+
+            Spacer()
+
+            if model.id == selectedModelId {
+                Image(systemName: "checkmark")
+                    .font(.exoSubheadline)
+                    .foregroundStyle(Color.exoYellow)
+            }
+        }
+    }
+}
@@ -0,0 +1,68 @@
+import SwiftUI
+
+struct ConnectionStatusBadge: View {
+    let connectionState: ConnectionState
+    var localModelState: LocalModelState = .notDownloaded
+
+    private var isLocalReady: Bool {
+        if case .disconnected = connectionState {
+            return localModelState == .ready || localModelState == .generating
+        }
+        return false
+    }
+
+    var body: some View {
+        HStack(spacing: 6) {
+            Circle()
+                .fill(dotColor)
+                .frame(width: 8, height: 8)
+                .shadow(color: dotColor.opacity(0.6), radius: 4)
+
+            Text(label.uppercased())
+                .font(.exoMono(10, weight: .medium))
+                .tracking(1)
+                .foregroundStyle(Color.exoForeground)
+        }
+        .padding(.horizontal, 10)
+        .padding(.vertical, 5)
+        .background(backgroundColor)
+        .clipShape(Capsule())
+        .overlay(
+            Capsule()
+                .stroke(dotColor.opacity(0.3), lineWidth: 1)
+        )
+    }
+
+    private var dotColor: Color {
+        if isLocalReady {
+            return .exoYellow
+        }
+        switch connectionState {
+        case .connected: return .green
+        case .connecting: return .orange
+        case .disconnected: return .exoLightGray
+        }
+    }
+
+    private var label: String {
+        if isLocalReady {
+            return "Local"
+        }
+        switch connectionState {
+        case .connected: return "Connected"
+        case .connecting: return "Connecting"
+        case .disconnected: return "Disconnected"
+        }
+    }
+
+    private var backgroundColor: Color {
+        if isLocalReady {
+            return Color.exoYellow.opacity(0.1)
+        }
+        switch connectionState {
+        case .connected: return .green.opacity(0.1)
+        case .connecting: return .orange.opacity(0.1)
+        case .disconnected: return Color.exoMediumGray.opacity(0.5)
+        }
+    }
+}
@@ -0,0 +1,136 @@
+import SwiftUI
+
+struct RootView: View {
+    @Environment(ClusterService.self) private var clusterService
+    @Environment(DiscoveryService.self) private var discoveryService
+    @Environment(ChatService.self) private var chatService
+    @Environment(LocalInferenceService.self) private var localInferenceService
+    @State private var showSettings = false
+    @State private var showConversations = false
+
+    var body: some View {
+        NavigationStack {
+            ChatView()
+                .navigationTitle("EXO")
+                .navigationBarTitleDisplayMode(.inline)
+                .toolbar {
+                    ToolbarItem(placement: .topBarLeading) {
+                        conversationMenuButton
+                    }
+
+                    ToolbarItem(placement: .principal) {
+                        ConnectionStatusBadge(
+                            connectionState: clusterService.connectionState,
+                            localModelState: localInferenceService.modelState
+                        )
+                    }
+
+                    ToolbarItem(placement: .topBarTrailing) {
+                        Button {
+                            showSettings = true
+                        } label: {
+                            Image(systemName: "gear")
+                                .foregroundStyle(Color.exoYellow)
+                        }
+                    }
+                }
+        }
+        .tint(Color.exoYellow)
+        .sheet(isPresented: $showSettings) {
+            SettingsView()
+                .environment(discoveryService)
+                .presentationBackground(Color.exoDarkGray)
+        }
+        .sheet(isPresented: $showConversations) {
+            conversationList
+                .presentationBackground(Color.exoDarkGray)
+        }
+    }
+
+    // MARK: - Conversations
+
+    private var conversationMenuButton: some View {
+        HStack(spacing: 12) {
+            Button {
+                showConversations = true
+            } label: {
+                Image(systemName: "sidebar.left")
+                    .foregroundStyle(Color.exoYellow)
+            }
+
+            Button {
+                chatService.createConversation()
+            } label: {
+                Image(systemName: "square.and.pencil")
+                    .foregroundStyle(Color.exoYellow)
+            }
+        }
+    }
+
+    private var conversationList: some View {
+        NavigationStack {
+            List {
+                if chatService.conversations.isEmpty {
+                    Text("No conversations yet")
+                        .font(.exoBody)
+                        .foregroundStyle(Color.exoLightGray)
+                        .listRowBackground(Color.exoDarkGray)
+                } else {
+                    ForEach(chatService.conversations) { conversation in
+                        let isActive = conversation.id == chatService.activeConversationId
+                        Button {
+                            chatService.setActiveConversation(id: conversation.id)
+                            showConversations = false
+                        } label: {
+                            VStack(alignment: .leading, spacing: 4) {
+                                Text(conversation.title)
+                                    .font(.exoSubheadline)
+                                    .fontWeight(isActive ? .semibold : .regular)
+                                    .foregroundStyle(
+                                        isActive ? Color.exoYellow : Color.exoForeground
+                                    )
+                                    .lineLimit(1)
+
+                                if let modelId = conversation.modelId {
+                                    Text(modelId)
+                                        .font(.exoCaption)
+                                        .foregroundStyle(Color.exoLightGray)
+                                        .lineLimit(1)
+                                }
+                            }
+                        }
+                        .listRowBackground(
+                            isActive
+                                ? Color.exoYellow.opacity(0.1)
+                                : Color.exoDarkGray
+                        )
+                    }
+                    .onDelete { indexSet in
+                        for index in indexSet {
+                            chatService.deleteConversation(id: chatService.conversations[index].id)
+                        }
+                    }
+                }
+            }
+            .scrollContentBackground(.hidden)
+            .background(Color.exoBlack)
+            .navigationTitle("Conversations")
+            .navigationBarTitleDisplayMode(.inline)
+            .toolbar {
+                ToolbarItem(placement: .confirmationAction) {
+                    Button("Done") { showConversations = false }
+                        .font(.exoSubheadline)
+                        .foregroundStyle(Color.exoYellow)
+                }
+                ToolbarItem(placement: .topBarLeading) {
+                    Button {
+                        chatService.createConversation()
+                    } label: {
+                        Image(systemName: "plus")
+                            .foregroundStyle(Color.exoYellow)
+                    }
+                }
+            }
+        }
+    }
+}
@@ -0,0 +1,314 @@
+import SwiftUI
+
+struct SettingsView: View {
+    @Environment(ClusterService.self) private var clusterService
+    @Environment(DiscoveryService.self) private var discoveryService
+    @Environment(LocalInferenceService.self) private var localInferenceService
+    @Environment(\.dismiss) private var dismiss
+    @State private var host: String = ""
+    @State private var port: String = "52415"
+
+    var body: some View {
+        NavigationStack {
+            Form {
+                localModelSection
+                nearbyClustersSection
+                connectionSection
+                statusSection
+            }
+            .scrollContentBackground(.hidden)
+            .background(Color.exoBlack)
+            .navigationTitle("Settings")
+            .navigationBarTitleDisplayMode(.inline)
+            .toolbar {
+                ToolbarItem(placement: .confirmationAction) {
+                    Button("Done") { dismiss() }
+                        .font(.exoSubheadline)
+                        .foregroundStyle(Color.exoYellow)
+                }
+            }
+        }
+    }
+
+    // MARK: - Section Headers
+
+    private func sectionHeader(_ title: String) -> some View {
+        Text(title.uppercased())
+            .font(.exoMono(10, weight: .semibold))
+            .tracking(2)
+            .foregroundStyle(Color.exoYellow)
+    }
+
+    // MARK: - Local Model
+
+    private var localModelSection: some View {
+        Section {
+            HStack {
+                VStack(alignment: .leading, spacing: 4) {
+                    Text(localInferenceService.defaultModelId)
+                        .font(.exoSubheadline)
+                        .foregroundStyle(Color.exoForeground)
+
+                    Text(localModelStatusText)
+                        .font(.exoCaption)
+                        .foregroundStyle(Color.exoLightGray)
+                }
+
+                Spacer()
+
+                localModelActionButton
+            }
+            .listRowBackground(Color.exoDarkGray)
+
+            if case .downloading(let progress) = localInferenceService.modelState {
+                ProgressView(value: progress)
+                    .tint(Color.exoYellow)
+                    .listRowBackground(Color.exoDarkGray)
+            }
+        } header: {
+            sectionHeader("Local Model")
+        } footer: {
+            Text(
+                "When disconnected from a cluster, messages are processed on-device using this model."
+            )
+            .font(.exoCaption)
+            .foregroundStyle(Color.exoLightGray.opacity(0.7))
+        }
+    }
+
+    private var localModelStatusText: String {
+        switch localInferenceService.modelState {
+        case .notDownloaded: "Not downloaded"
+        case .downloading(let progress): "Downloading \(Int(progress * 100))%..."
+        case .downloaded: "Downloaded — not loaded"
+        case .loading: "Loading into memory..."
+        case .ready: "Ready"
+        case .generating: "Generating..."
+        case .error(let message): "Error: \(message)"
+        }
+    }
+
+    @ViewBuilder
+    private var localModelActionButton: some View {
+        switch localInferenceService.modelState {
+        case .notDownloaded:
+            exoButton("Download") {
+                Task { await localInferenceService.prepareModel() }
+            }
+        case .downloading:
+            ProgressView()
+                .controlSize(.small)
+                .tint(Color.exoYellow)
+        case .downloaded:
+            exoButton("Load") {
+                Task { await localInferenceService.prepareModel() }
+            }
+        case .loading:
+            ProgressView()
+                .controlSize(.small)
+                .tint(Color.exoYellow)
+        case .ready, .generating:
+            exoButton("Unload") {
+                localInferenceService.unloadModel()
+            }
+        case .error:
+            exoButton("Retry", destructive: true) {
+                Task { await localInferenceService.prepareModel() }
+            }
+        }
+    }
+
+    private func exoButton(_ title: String, destructive: Bool = false, action: @escaping () -> Void)
+        -> some View
+    {
+        let borderColor = destructive ? Color.exoDestructive : Color.exoYellow
+        return Button(action: action) {
+            Text(title.uppercased())
+                .font(.exoMono(11, weight: .semibold))
+                .tracking(1)
+                .foregroundStyle(borderColor)
+                .padding(.horizontal, 10)
+                .padding(.vertical, 5)
+                .overlay(
+                    RoundedRectangle(cornerRadius: 6)
+                        .stroke(borderColor, lineWidth: 1)
+                )
+        }
+    }
+
+    // MARK: - Nearby Clusters
+
+    private var nearbyClustersSection: some View {
+        Section {
+            if discoveryService.discoveredClusters.isEmpty {
+                if discoveryService.isSearching {
+                    HStack {
+                        ProgressView()
+                            .tint(Color.exoYellow)
+                            .padding(.trailing, 8)
+                        Text("Searching for clusters...")
+                            .font(.exoBody)
+                            .foregroundStyle(Color.exoLightGray)
+                    }
+                    .listRowBackground(Color.exoDarkGray)
+                } else {
+                    Text("No clusters found")
+                        .font(.exoBody)
+                        .foregroundStyle(Color.exoLightGray)
+                        .listRowBackground(Color.exoDarkGray)
+                }
+            } else {
+                ForEach(discoveryService.discoveredClusters) { cluster in
+                    HStack {
+                        VStack(alignment: .leading) {
+                            Text(cluster.name)
+                                .font(.exoBody)
+                                .foregroundStyle(Color.exoForeground)
+                        }
+                        Spacer()
+                        exoButton("Connect") {
+                            Task {
+                                await clusterService.connectToDiscoveredCluster(
+                                    cluster, using: discoveryService
+                                )
+                                if clusterService.isConnected {
+                                    dismiss()
+                                }
+                            }
+                        }
+                    }
+                    .listRowBackground(Color.exoDarkGray)
+                }
+            }
+        } header: {
+            sectionHeader("Nearby Clusters")
+        }
+    }
+
+    // MARK: - Manual Connection
+
+    private var connectionSection: some View {
+        Section {
+            TextField("IP Address (e.g. 192.168.1.42)", text: $host)
+                .font(.exoBody)
+                .keyboardType(.decimalPad)
+                .textContentType(.URL)
+                .autocorrectionDisabled()
+                .foregroundStyle(Color.exoForeground)
+                .listRowBackground(Color.exoDarkGray)
+
+            TextField("Port", text: $port)
+                .font(.exoBody)
+                .keyboardType(.numberPad)
+                .foregroundStyle(Color.exoForeground)
+                .listRowBackground(Color.exoDarkGray)
+
+            Button {
+                Task {
+                    let portNum = Int(port) ?? ConnectionInfo.defaultPort
+                    let info = ConnectionInfo(host: host, port: portNum, nodeId: nil)
+                    await clusterService.connect(to: info)
+                    if clusterService.isConnected {
+                        dismiss()
+                    }
+                }
+            } label: {
+                Text(clusterService.isConnected ? "RECONNECT" : "CONNECT")
+                    .font(.exoMono(13, weight: .semibold))
+                    .tracking(1.5)
+                    .foregroundStyle(
+                        host.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
+                            ? Color.exoLightGray : Color.exoYellow
+                    )
+                    .frame(maxWidth: .infinity)
+            }
+            .disabled(host.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty)
+            .listRowBackground(Color.exoDarkGray)
+        } header: {
+            sectionHeader("Manual Connection")
+        }
+    }
+
+    // MARK: - Status
+
+    private var statusSection: some View {
+        Section {
+            if let connection = clusterService.currentConnection {
+                LabeledContent {
+                    Text(connection.host)
+                        .font(.exoCaption)
+                        .foregroundStyle(Color.exoForeground)
+                } label: {
+                    Text("Host")
+                        .font(.exoCaption)
+                        .foregroundStyle(Color.exoLightGray)
+                }
+                .listRowBackground(Color.exoDarkGray)
+
+                LabeledContent {
+                    Text("\(connection.port)")
+                        .font(.exoCaption)
+                        .foregroundStyle(Color.exoForeground)
+                } label: {
+                    Text("Port")
+                        .font(.exoCaption)
+                        .foregroundStyle(Color.exoLightGray)
+                }
+                .listRowBackground(Color.exoDarkGray)
+
+                if let nodeId = connection.nodeId {
+                    LabeledContent {
+                        Text(String(nodeId.prefix(12)) + "...")
+                            .font(.exoCaption)
+                            .foregroundStyle(Color.exoForeground)
+                    } label: {
+                        Text("Node ID")
+                            .font(.exoCaption)
+                            .foregroundStyle(Color.exoLightGray)
+                    }
+                    .listRowBackground(Color.exoDarkGray)
+                }
+
+                LabeledContent {
+                    Text("\(clusterService.availableModels.count)")
+                        .font(.exoCaption)
+                        .foregroundStyle(Color.exoForeground)
+                } label: {
+                    Text("Models")
+                        .font(.exoCaption)
+                        .foregroundStyle(Color.exoLightGray)
+                }
+                .listRowBackground(Color.exoDarkGray)
+
+                Button(role: .destructive) {
+                    clusterService.disconnect()
+                } label: {
+                    Text("DISCONNECT")
+                        .font(.exoMono(13, weight: .semibold))
+                        .tracking(1.5)
+                        .foregroundStyle(Color.exoDestructive)
+                        .frame(maxWidth: .infinity)
+                }
+                .listRowBackground(Color.exoDarkGray)
+            } else {
+                if let error = clusterService.lastError {
+                    Label {
+                        Text(error)
+                            .font(.exoCaption)
+                    } icon: {
+                        Image(systemName: "exclamationmark.triangle")
+                    }
+                    .foregroundStyle(Color.exoDestructive)
+                    .listRowBackground(Color.exoDarkGray)
+                } else {
+                    Text("Not connected")
+                        .font(.exoBody)
+                        .foregroundStyle(Color.exoLightGray)
+                        .listRowBackground(Color.exoDarkGray)
+                }
+            }
+        } header: {
+            sectionHeader("Status")
+        }
+    }
+}
@@ -0,0 +1,51 @@
+import SwiftUI
+
+// MARK: - EXO Color Palette
+
+extension Color {
+    /// Primary background — near-black (#121212)
+    static let exoBlack = Color(red: 0x12 / 255.0, green: 0x12 / 255.0, blue: 0x12 / 255.0)
+    /// Card / surface background (#1F1F1F)
+    static let exoDarkGray = Color(red: 0x1F / 255.0, green: 0x1F / 255.0, blue: 0x1F / 255.0)
+    /// Input field / elevated surface (#353535)
+    static let exoMediumGray = Color(red: 0x35 / 255.0, green: 0x35 / 255.0, blue: 0x35 / 255.0)
+    /// Secondary text (#999999)
+    static let exoLightGray = Color(red: 0x99 / 255.0, green: 0x99 / 255.0, blue: 0x99 / 255.0)
+    /// Accent yellow — matches dashboard (#FFD700)
+    static let exoYellow = Color(red: 0xFF / 255.0, green: 0xD7 / 255.0, blue: 0x00 / 255.0)
+    /// Primary foreground text (#E5E5E5)
+    static let exoForeground = Color(red: 0xE5 / 255.0, green: 0xE5 / 255.0, blue: 0xE5 / 255.0)
+    /// Destructive / error (#E74C3C)
+    static let exoDestructive = Color(red: 0xE7 / 255.0, green: 0x4C / 255.0, blue: 0x3C / 255.0)
+}
+
+// MARK: - EXO Typography (SF Mono via .monospaced design)
+
+extension Font {
+    /// Monospaced font at a given size and weight.
+    static func exoMono(_ size: CGFloat, weight: Font.Weight = .regular) -> Font {
+        .system(size: size, weight: weight, design: .monospaced)
+    }
+
+    /// Body text — 15pt monospaced
+    static let exoBody: Font = .system(size: 15, weight: .regular, design: .monospaced)
+    /// Caption — 11pt monospaced
+    static let exoCaption: Font = .system(size: 11, weight: .regular, design: .monospaced)
+    /// Subheadline — 13pt monospaced medium
+    static let exoSubheadline: Font = .system(size: 13, weight: .medium, design: .monospaced)
+    /// Headline — 17pt monospaced semibold
+    static let exoHeadline: Font = .system(size: 17, weight: .semibold, design: .monospaced)
+}
+
+// MARK: - Reusable Gradient Divider
+
+struct GradientDivider: View {
+    var body: some View {
+        LinearGradient(
+            colors: [.clear, Color.exoYellow.opacity(0.3), .clear],
+            startPoint: .leading,
+            endPoint: .trailing
+        )
+        .frame(height: 1)
+    }
+}
@@ -0,0 +1,18 @@
+//
+//  EXO_iOSTests.swift
+//  EXO-iOSTests
+//
+//  Created by Sami Khan on 2026-02-17.
+//
+
+import Testing
+
+@testable import EXO_iOS
+
+struct EXO_iOSTests {
+
+    @Test func example() async throws {
+        // Write your test here and use APIs like `#expect(...)` to check expected conditions.
+    }
+
+}
@@ -0,0 +1,41 @@
+//
+//  EXO_iOSUITests.swift
+//  EXO-iOSUITests
+//
+//  Created by Sami Khan on 2026-02-17.
+//
+
+import XCTest
+
+final class EXO_iOSUITests: XCTestCase {
+
+    override func setUpWithError() throws {
+        // Put setup code here. This method is called before the invocation of each test method in the class.
+
+        // In UI tests it is usually best to stop immediately when a failure occurs.
+        continueAfterFailure = false
+
+        // In UI tests it’s important to set the initial state - such as interface orientation - required for your tests before they run. The setUp method is a good place to do this.
+    }
+
+    override func tearDownWithError() throws {
+        // Put teardown code here. This method is called after the invocation of each test method in the class.
+    }
+
+    @MainActor
+    func testExample() throws {
+        // UI tests must launch the application that they test.
+        let app = XCUIApplication()
+        app.launch()
+
+        // Use XCTAssert and related functions to verify your tests produce the correct results.
+    }
+
+    @MainActor
+    func testLaunchPerformance() throws {
+        // This measures how long it takes to launch your application.
+        measure(metrics: [XCTApplicationLaunchMetric()]) {
+            XCUIApplication().launch()
+        }
+    }
+}
@@ -0,0 +1,33 @@
+//
+//  EXO_iOSUITestsLaunchTests.swift
+//  EXO-iOSUITests
+//
+//  Created by Sami Khan on 2026-02-17.
+//
+
+import XCTest
+
+final class EXO_iOSUITestsLaunchTests: XCTestCase {
+
+    override class var runsForEachTargetApplicationUIConfiguration: Bool {
+        true
+    }
+
+    override func setUpWithError() throws {
+        continueAfterFailure = false
+    }
+
+    @MainActor
+    func testLaunch() throws {
+        let app = XCUIApplication()
+        app.launch()
+
+        // Insert steps here to perform after app launch but before taking a screenshot,
+        // such as logging into a test account or navigating somewhere in the app
+
+        let attachment = XCTAttachment(screenshot: app.screenshot())
+        attachment.name = "Launch Screen"
+        attachment.lifetime = .keepAlways
+        add(attachment)
+    }
+}
@@ -563,21 +563,45 @@ struct ContentView: View {
    }

    private var rdmaStatusView: some View {
-        let rdma = networkStatusService.status.rdmaStatus
+        let rdmaStatuses = stateService.latestSnapshot?.nodeRdmaCtl ?? [:]
+        let localNodeId = stateService.localNodeId
+        let nodeProfiles = stateService.latestSnapshot?.nodeProfiles ?? [:]
+        let localDevices = networkStatusService.status.localRdmaDevices
+        let localPorts = networkStatusService.status.localRdmaActivePorts
+
        return VStack(alignment: .leading, spacing: 1) {
-            Text("RDMA: \(rdmaStatusText(rdma))")
-                .font(.caption2)
-                .foregroundColor(rdmaStatusColor(rdma))
-            if !rdma.devices.isEmpty {
-                Text("  Devices: \(rdma.devices.joined(separator: ", "))")
+            if rdmaStatuses.isEmpty {
+                Text("Cluster RDMA: No data")
+                    .font(.caption2)
+                    .foregroundColor(.secondary)
+            } else {
+                Text("Cluster RDMA Status:")
+                    .font(.caption2)
+                    .foregroundColor(.secondary)
+                ForEach(Array(rdmaStatuses.keys.sorted()), id: \.self) { nodeId in
+                    if let status = rdmaStatuses[nodeId] {
+                        let nodeName =
+                            nodeProfiles[nodeId]?.friendlyName ?? String(nodeId.prefix(8))
+                        let isLocal = nodeId == localNodeId
+                        let prefix = isLocal ? "  \(nodeName) (local):" : "  \(nodeName):"
+                        let statusText = status.enabled ? "Enabled" : "Disabled"
+                        let color: Color = status.enabled ? .green : .orange
+                        Text("\(prefix) \(statusText)")
+                            .font(.caption2)
+                            .foregroundColor(color)
+                    }
+                }
+            }
+            if !localDevices.isEmpty {
+                Text("  Local Devices: \(localDevices.joined(separator: ", "))")
                    .font(.caption2)
                    .foregroundColor(.secondary)
            }
-            if !rdma.activePorts.isEmpty {
-                Text("  Active Ports:")
+            if !localPorts.isEmpty {
+                Text("  Local Active Ports:")
                    .font(.caption2)
                    .foregroundColor(.secondary)
-                ForEach(rdma.activePorts, id: \.device) { port in
+                ForEach(localPorts, id: \.device) { port in
                    Text("    \(port.device) port \(port.port): \(port.state)")
                        .font(.caption2)
                        .foregroundColor(.green)
@@ -586,28 +610,6 @@ struct ContentView: View {
        }
    }

-    private func rdmaStatusText(_ rdma: RDMAStatus) -> String {
-        switch rdma.rdmaCtlEnabled {
-        case .some(true):
-            return "Enabled"
-        case .some(false):
-            return "Disabled"
-        case nil:
-            return rdma.devices.isEmpty ? "Not Available" : "Available"
-        }
-    }
-
-    private func rdmaStatusColor(_ rdma: RDMAStatus) -> Color {
-        switch rdma.rdmaCtlEnabled {
-        case .some(true):
-            return .green
-        case .some(false):
-            return .orange
-        case nil:
-            return rdma.devices.isEmpty ? .secondary : .green
-        }
-    }
-
    private var sendBugReportButton: some View {
        VStack(alignment: .leading, spacing: 4) {
            Button {
@@ -28,6 +28,10 @@ final class ExoProcessController: ObservableObject {
        }
    }

+    static let exoDirectoryURL: URL = {
+        URL(fileURLWithPath: NSHomeDirectory()).appendingPathComponent(".exo")
+    }()
+
    @Published private(set) var status: Status = .stopped
    @Published private(set) var lastError: String?
    @Published private(set) var launchCountdownSeconds: Int?
@@ -78,7 +82,11 @@ final class ExoProcessController: ObservableObject {

            let child = Process()
            child.executableURL = executableURL
-            child.currentDirectoryURL = runtimeURL
+            let exoHomeURL = Self.exoDirectoryURL
+            try? FileManager.default.createDirectory(
+                at: exoHomeURL, withIntermediateDirectories: true
+            )
+            child.currentDirectoryURL = exoHomeURL
            child.environment = makeEnvironment(for: runtimeURL)

            child.standardOutput = FileHandle.nullDevice
@@ -15,6 +15,7 @@ struct ClusterState: Decodable {
    let nodeMemory: [String: MemoryInfo]
    let nodeSystem: [String: SystemInfo]
    let nodeThunderboltBridge: [String: ThunderboltBridgeStatus]
+    let nodeRdmaCtl: [String: NodeRdmaCtlStatus]

    /// Computed property for backwards compatibility - merges granular state into NodeProfile
    var nodeProfiles: [String: NodeProfile] {
@@ -65,6 +66,10 @@ struct ClusterState: Decodable {
            try container.decodeIfPresent(
                [String: ThunderboltBridgeStatus].self, forKey: .nodeThunderboltBridge
            ) ?? [:]
+        self.nodeRdmaCtl =
+            try container.decodeIfPresent(
+                [String: NodeRdmaCtlStatus].self, forKey: .nodeRdmaCtl
+            ) ?? [:]
    }

    private enum CodingKeys: String, CodingKey {
@@ -78,6 +83,7 @@ struct ClusterState: Decodable {
        case nodeMemory
        case nodeSystem
        case nodeThunderboltBridge
+        case nodeRdmaCtl
    }
 }

@@ -159,6 +165,10 @@ struct ThunderboltBridgeStatus: Decodable {
    let serviceName: String?
 }

+struct NodeRdmaCtlStatus: Decodable {
+    let enabled: Bool
+}
+
 struct MemoryInfo: Decodable {
    let ramTotal: MemoryValue?
    let ramAvailable: MemoryValue?
@@ -44,16 +44,12 @@ struct BugReportService {
        let dayPrefix = Self.dayPrefixString(now)
        let prefix = "reports/\(dayPrefix)/\(timestamp)/"

-        let logData = readLog()
+        let logFiles = readAllLogs()
        let ifconfigText = try await captureIfconfig()
        let hostName = Host.current().localizedName ?? "unknown"
        let debugInfo = readDebugInfo()

-        async let stateResult = fetch(url: baseURL.appendingPathComponent("state"))
-        async let eventsResult = fetch(url: baseURL.appendingPathComponent("events"))
-
-        let stateData = try await stateResult
-        let eventsData = try await eventsResult
+        let stateData = try await fetch(url: baseURL.appendingPathComponent("state"))

        // Extract cluster TB bridge status from exo state
        let clusterTbBridgeStatus = extractClusterTbBridgeStatus(from: stateData)
@@ -67,12 +63,19 @@ struct BugReportService {
            clusterTbBridgeStatus: clusterTbBridgeStatus
        )

-        let uploads: [(path: String, data: Data?)] = [
-            ("\(prefix)exo.log", logData),
+        let eventLogFiles = readAllEventLogs()
+
+        var uploads: [(path: String, data: Data?)] = logFiles.map { (path, data) in
+            ("\(prefix)\(path)", data)
+        }
+        uploads.append(
+            contentsOf: eventLogFiles.map { (path, data) in
+                ("\(prefix)\(path)", data as Data?)
+            })
+        uploads.append(contentsOf: [
            ("\(prefix)state.json", stateData),
-            ("\(prefix)events.json", eventsData),
            ("\(prefix)report.json", reportJSON),
-        ]
+        ])

        let uploadItems: [(key: String, body: Data)] = uploads.compactMap { item in
            guard let body = item.data else { return nil }
@@ -149,11 +152,40 @@ struct BugReportService {
        return decoded.urls
    }

-    private func readLog() -> Data? {
-        let logURL = URL(fileURLWithPath: NSHomeDirectory())
+    private func readAllLogs() -> [(path: String, data: Data)] {
+        let dir = URL(fileURLWithPath: NSHomeDirectory())
            .appendingPathComponent(".exo")
-            .appendingPathComponent("exo.log")
-        return try? Data(contentsOf: logURL)
+            .appendingPathComponent("exo_log")
+        var results: [(path: String, data: Data)] = []
+
+        let contents = (try? FileManager.default.contentsOfDirectory(atPath: dir.path)) ?? []
+        for name in contents {
+            if let data = try? Data(contentsOf: dir.appendingPathComponent(name)) {
+                results.append(("exo_log/\(name)", data))
+            }
+        }
+
+        return results
+    }
+
+    private func readAllEventLogs() -> [(path: String, data: Data)] {
+        let eventLogDir = URL(fileURLWithPath: NSHomeDirectory())
+            .appendingPathComponent(".exo")
+            .appendingPathComponent("event_log")
+        var results: [(path: String, data: Data)] = []
+
+        for subdir in ["master", "api"] {
+            let dir = eventLogDir.appendingPathComponent(subdir)
+            let contents =
+                (try? FileManager.default.contentsOfDirectory(atPath: dir.path)) ?? []
+            for name in contents where name.hasPrefix("events.") {
+                if let data = try? Data(contentsOf: dir.appendingPathComponent(name)) {
+                    results.append(("event_log/\(subdir)/\(name)", data))
+                }
+            }
+        }
+
+        return results
    }

    private func captureIfconfig() async throws -> String {
@@ -35,28 +35,18 @@ struct NetworkStatus: Equatable {
    let thunderboltBridgeState: ThunderboltState?
    let bridgeInactive: Bool?
    let interfaceStatuses: [InterfaceIpStatus]
-    let rdmaStatus: RDMAStatus
+    let localRdmaDevices: [String]
+    let localRdmaActivePorts: [RDMAPort]

    static let empty = NetworkStatus(
        thunderboltBridgeState: nil,
        bridgeInactive: nil,
        interfaceStatuses: [],
-        rdmaStatus: .empty
+        localRdmaDevices: [],
+        localRdmaActivePorts: []
    )
 }

-struct RDMAStatus: Equatable {
-    let rdmaCtlEnabled: Bool?
-    let devices: [String]
-    let activePorts: [RDMAPort]
-
-    var isAvailable: Bool {
-        rdmaCtlEnabled == true || !devices.isEmpty
-    }
-
-    static let empty = RDMAStatus(rdmaCtlEnabled: nil, devices: [], activePorts: [])
-}
-
 struct RDMAPort: Equatable {
    let device: String
    let port: String
@@ -80,31 +70,11 @@ private struct NetworkStatusFetcher {
            thunderboltBridgeState: readThunderboltBridgeState(),
            bridgeInactive: readBridgeInactive(),
            interfaceStatuses: readInterfaceStatuses(),
-            rdmaStatus: readRDMAStatus()
+            localRdmaDevices: readRDMADevices(),
+            localRdmaActivePorts: readRDMAActivePorts()
        )
    }

-    private func readRDMAStatus() -> RDMAStatus {
-        let rdmaCtlEnabled = readRDMACtlEnabled()
-        let devices = readRDMADevices()
-        let activePorts = readRDMAActivePorts()
-        return RDMAStatus(
-            rdmaCtlEnabled: rdmaCtlEnabled, devices: devices, activePorts: activePorts)
-    }
-
-    private func readRDMACtlEnabled() -> Bool? {
-        let result = runCommand(["rdma_ctl", "status"])
-        guard result.exitCode == 0 else { return nil }
-        let output = result.output.lowercased().trimmingCharacters(in: .whitespacesAndNewlines)
-        if output.contains("enabled") {
-            return true
-        }
-        if output.contains("disabled") {
-            return false
-        }
-        return nil
-    }
-
    private func readRDMADevices() -> [String] {
        let result = runCommand(["ibv_devices"])
        guard result.exitCode == 0 else { return [] }
@@ -0,0 +1,7 @@
+# Canary benchmark manifest
+#
+# Lists the suite files to include. Each file defines benchmarks
+# with shared constraints, topology, and default args.
+include = [
+    "single-m3-ultra.toml",
+]
@@ -19,6 +19,11 @@ from urllib.parse import urlencode
 from loguru import logger
 from transformers import AutoTokenizer

+# Backoff constants for cluster settling retry
+_SETTLE_INITIAL_BACKOFF_S = 1.0
+_SETTLE_MAX_BACKOFF_S = 60.0
+_SETTLE_BACKOFF_MULTIPLIER = 2.0
+
 # Monkey-patch for transformers 5.x compatibility
 # Kimi's tokenization_kimi.py imports bytes_to_unicode from the old location
 # which was moved in transformers 5.0.0rc2
@@ -283,6 +288,151 @@ def resolve_model_short_id(client: ExoClient, model_arg: str) -> tuple[str, str]
    raise ValueError(f"Model not found in /models: {model_arg}")


+def run_planning_phase(
+    client: ExoClient,
+    full_model_id: str,
+    preview: dict[str, Any],
+    danger_delete: bool,
+    timeout: float,
+    settle_deadline: float | None,
+) -> None:
+    """Check disk space and ensure model is downloaded before benchmarking."""
+    # Get model size from /models
+    models = client.request_json("GET", "/models") or {}
+    model_bytes = 0
+    for m in models.get("data", []):
+        if m.get("hugging_face_id") == full_model_id:
+            model_bytes = m.get("storage_size_megabytes", 0) * 1024 * 1024
+            break
+
+    if not model_bytes:
+        logger.warning(
+            f"Could not determine size for {full_model_id}, skipping disk check"
+        )
+        return
+
+    # Get nodes from preview
+    inner = unwrap_instance(preview["instance"])
+    node_ids = list(inner["shardAssignments"]["nodeToRunner"].keys())
+    runner_to_shard = inner["shardAssignments"]["runnerToShard"]
+
+    state = client.request_json("GET", "/state")
+    downloads = state.get("downloads", {})
+    node_disk = state.get("nodeDisk", {})
+
+    for node_id in node_ids:
+        node_downloads = downloads.get(node_id, [])
+
+        # Check if model already downloaded on this node
+        already_downloaded = any(
+            "DownloadCompleted" in p
+            and unwrap_instance(p["DownloadCompleted"]["shardMetadata"])["modelCard"][
+                "modelId"
+            ]
+            == full_model_id
+            for p in node_downloads
+        )
+        if already_downloaded:
+            continue
+
+        # Wait for disk info if settle_deadline is set
+        disk_info = node_disk.get(node_id, {})
+        backoff = _SETTLE_INITIAL_BACKOFF_S
+        while not disk_info and settle_deadline and time.monotonic() < settle_deadline:
+            remaining = settle_deadline - time.monotonic()
+            logger.info(
+                f"Waiting for disk info on {node_id} ({remaining:.0f}s remaining)..."
+            )
+            time.sleep(min(backoff, remaining))
+            backoff = min(backoff * _SETTLE_BACKOFF_MULTIPLIER, _SETTLE_MAX_BACKOFF_S)
+            state = client.request_json("GET", "/state")
+            node_disk = state.get("nodeDisk", {})
+            disk_info = node_disk.get(node_id, {})
+
+        if not disk_info:
+            logger.warning(f"No disk info for {node_id}, skipping space check")
+            continue
+
+        avail = disk_info.get("available", {}).get("inBytes", 0)
+        if avail >= model_bytes:
+            continue
+
+        if not danger_delete:
+            raise RuntimeError(
+                f"Insufficient disk on {node_id}: need {model_bytes // (1024**3)}GB, "
+                f"have {avail // (1024**3)}GB. Use --danger-delete-downloads to free space."
+            )
+
+        # Delete from smallest to largest
+        completed = [
+            (
+                unwrap_instance(p["DownloadCompleted"]["shardMetadata"])["modelCard"][
+                    "modelId"
+                ],
+                p["DownloadCompleted"]["totalBytes"]["inBytes"],
+            )
+            for p in node_downloads
+            if "DownloadCompleted" in p
+        ]
+        for del_model, size in sorted(completed, key=lambda x: x[1]):
+            logger.info(f"Deleting {del_model} from {node_id} ({size // (1024**2)}MB)")
+            client.request_json("DELETE", f"/download/{node_id}/{del_model}")
+            avail += size
+            if avail >= model_bytes:
+                break
+
+        if avail < model_bytes:
+            raise RuntimeError(f"Could not free enough space on {node_id}")
+
+    # Start downloads (idempotent)
+    for node_id in node_ids:
+        runner_id = inner["shardAssignments"]["nodeToRunner"][node_id]
+        shard = runner_to_shard[runner_id]
+        client.request_json(
+            "POST",
+            "/download/start",
+            body={
+                "targetNodeId": node_id,
+                "shardMetadata": shard,
+            },
+        )
+        logger.info(f"Started download on {node_id}")
+
+    # Wait for downloads
+    start = time.time()
+    while time.time() - start < timeout:
+        state = client.request_json("GET", "/state")
+        downloads = state.get("downloads", {})
+        all_done = True
+        for node_id in node_ids:
+            done = any(
+                "DownloadCompleted" in p
+                and unwrap_instance(p["DownloadCompleted"]["shardMetadata"])[
+                    "modelCard"
+                ]["modelId"]
+                == full_model_id
+                for p in downloads.get(node_id, [])
+            )
+            failed = [
+                p["DownloadFailed"]["errorMessage"]
+                for p in downloads.get(node_id, [])
+                if "DownloadFailed" in p
+                and unwrap_instance(p["DownloadFailed"]["shardMetadata"])["modelCard"][
+                    "modelId"
+                ]
+                == full_model_id
+            ]
+            if failed:
+                raise RuntimeError(f"Download failed on {node_id}: {failed[0]}")
+            if not done:
+                all_done = False
+        if all_done:
+            return
+        time.sleep(1)
+
+    raise TimeoutError("Downloads did not complete in time")
+
+
 def placement_filter(instance_meta: str, wanted: str) -> bool:
    s = (instance_meta or "").lower()
    if wanted == "both":
@@ -388,6 +538,66 @@ class PromptSizer:
        return content, tok


+def fetch_and_filter_placements(
+    client: ExoClient, full_model_id: str, args: argparse.Namespace
+) -> list[dict[str, Any]]:
+    previews_resp = client.request_json(
+        "GET", "/instance/previews", params={"model_id": full_model_id}
+    )
+    previews = previews_resp.get("previews") or []
+
+    selected: list[dict[str, Any]] = []
+    for p in previews:
+        if p.get("error") is not None:
+            continue
+        if not placement_filter(str(p.get("instance_meta", "")), args.instance_meta):
+            continue
+        if not sharding_filter(str(p.get("sharding", "")), args.sharding):
+            continue
+
+        instance = p.get("instance")
+        if not isinstance(instance, dict):
+            continue
+
+        n = nodes_used_in_instance(instance)
+        # Skip tensor ring single node as it is pointless when pipeline ring
+        if n == 1 and (
+            (args.sharding == "both" and "tensor" in p.get("sharding", "").lower())
+            or (
+                args.instance_meta == "both"
+                and "jaccl" in p.get("instance_meta", "").lower()
+            )
+        ):
+            continue
+
+        if (
+            args.skip_pipeline_jaccl
+            and (
+                args.instance_meta == "both"
+                and "jaccl" in p.get("instance_meta", "").lower()
+            )
+            and (
+                args.sharding == "both" and "pipeline" in p.get("sharding", "").lower()
+            )
+        ):
+            continue
+
+        if (
+            args.skip_tensor_ring
+            and (
+                args.instance_meta == "both"
+                and "ring" in p.get("instance_meta", "").lower()
+            )
+            and (args.sharding == "both" and "tensor" in p.get("sharding", "").lower())
+        ):
+            continue
+
+        if args.min_nodes <= n <= args.max_nodes:
+            selected.append(p)
+
+    return selected
+
+
 def main() -> int:
    ap = argparse.ArgumentParser(
        prog="exo-bench",
@@ -464,6 +674,17 @@ def main() -> int:
        action="store_true",
        help="Force all pp×tg combinations (cartesian product) even when lists have equal length.",
    )
+    ap.add_argument(
+        "--settle-timeout",
+        type=float,
+        default=0,
+        help="Max seconds to wait for the cluster to produce valid placements (0 = try once).",
+    )
+    ap.add_argument(
+        "--danger-delete-downloads",
+        action="store_true",
+        help="Delete existing models from smallest to largest to make room for benchmark model.",
+    )
    args = ap.parse_args()

    pp_list = parse_int_list(args.pp)
@@ -487,11 +708,6 @@ def main() -> int:
    client = ExoClient(args.host, args.port, timeout_s=args.timeout)
    short_id, full_model_id = resolve_model_short_id(client, args.model)

-    previews_resp = client.request_json(
-        "GET", "/instance/previews", params={"model_id": full_model_id}
-    )
-    previews = previews_resp.get("previews") or []
-
    tokenizer = load_tokenizer_for_bench(full_model_id)
    if tokenizer is None:
        raise RuntimeError("[exo-bench] tokenizer load failed")
@@ -503,54 +719,23 @@ def main() -> int:
        logger.error("[exo-bench] tokenizer usable but prompt sizing failed")
        raise

-    selected: list[dict[str, Any]] = []
-    for p in previews:
-        if p.get("error") is not None:
-            continue
-        if not placement_filter(str(p.get("instance_meta", "")), args.instance_meta):
-            continue
-        if not sharding_filter(str(p.get("sharding", "")), args.sharding):
-            continue
+    settle_deadline = (
+        time.monotonic() + args.settle_timeout if args.settle_timeout > 0 else None
+    )

-        instance = p.get("instance")
-        if not isinstance(instance, dict):
-            continue
+    selected = fetch_and_filter_placements(client, full_model_id, args)

-        n = nodes_used_in_instance(instance)
-        # Skip tensor ring single node as it is pointless when pipeline ring
-        if n == 1 and (
-            (args.sharding == "both" and "tensor" in p.get("sharding", "").lower())
-            or (
-                args.instance_meta == "both"
-                and "jaccl" in p.get("instance_meta", "").lower()
+    if not selected and settle_deadline:
+        backoff = _SETTLE_INITIAL_BACKOFF_S
+        while not selected and time.monotonic() < settle_deadline:
+            remaining = settle_deadline - time.monotonic()
+            logger.warning(
+                f"No valid placements yet (cluster may still be settling). "
+                f"Retrying in {backoff:.1f}s ({remaining:.0f}s remaining)..."
            )
-        ):
-            continue
-
-        if (
-            args.skip_pipeline_jaccl
-            and (
-                args.instance_meta == "both"
-                and "jaccl" in p.get("instance_meta", "").lower()
-            )
-            and (
-                args.sharding == "both" and "pipeline" in p.get("sharding", "").lower()
-            )
-        ):
-            continue
-
-        if (
-            args.skip_tensor_ring
-            and (
-                args.instance_meta == "both"
-                and "ring" in p.get("instance_meta", "").lower()
-            )
-            and (args.sharding == "both" and "tensor" in p.get("sharding", "").lower())
-        ):
-            continue
-
-        if args.min_nodes <= n <= args.max_nodes:
-            selected.append(p)
+            time.sleep(min(backoff, remaining))
+            backoff = min(backoff * _SETTLE_BACKOFF_MULTIPLIER, _SETTLE_MAX_BACKOFF_S)
+            selected = fetch_and_filter_placements(client, full_model_id, args)

    if not selected:
        logger.error("No valid placements matched your filters.")
@@ -575,6 +760,16 @@ def main() -> int:
    if args.dry_run:
        return 0

+    logger.info("Planning phase: checking downloads...")
+    run_planning_phase(
+        client,
+        full_model_id,
+        selected[0],
+        args.danger_delete_downloads,
+        args.timeout,
+        settle_deadline,
+    )
+
    all_rows: list[dict[str, Any]] = []

    for preview in selected:
@@ -0,0 +1,189 @@
+# Single-node M3 Ultra benchmarks
+#
+# Shared constraints applied to ALL benchmarks in this file.
+constraints = [
+    "All(MacOsBuild(=25D125))",
+    "Hosts(=1)",
+    "All(Chip(m3_ultra))",
+    "All(GpuCores(=80))",
+]
+
+[topology]
+type = "none"
+
+# Default args merged into each benchmark's args (benchmark-level args win).
+[defaults]
+pp = [512, 2048, 8192, 16384]
+tg = 128
+
+[[benchmark]]
+model = "mlx-community/Meta-Llama-3.1-70B-Instruct-4bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/gpt-oss-120b-MXFP4-Q8"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/GLM-4.7-Flash-8bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-Coder-Next-6bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-30B-A3B-8bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-0.6B-4bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-0.6B-8bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Llama-3.2-1B-Instruct-4bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Llama-3.2-3B-Instruct-4bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Llama-3.2-3B-Instruct-8bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Meta-Llama-3.1-8B-Instruct-4bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Meta-Llama-3.1-8B-Instruct-8bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Meta-Llama-3.1-8B-Instruct-bf16"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/gpt-oss-20b-MXFP4-Q8"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-30B-A3B-4bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/GLM-4.7-Flash-4bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/GLM-4.7-Flash-5bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/GLM-4.7-Flash-6bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Llama-3.3-70B-Instruct-4bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-Coder-Next-4bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-Coder-Next-5bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-Coder-Next-8bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-Next-80B-A3B-Instruct-4bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-Next-80B-A3B-Instruct-8bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-Next-80B-A3B-Thinking-4bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-Next-80B-A3B-Thinking-8bit"
+extra_constraints = ["All(Memory(>=96GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Llama-3.3-70B-Instruct-8bit"
+extra_constraints = ["All(Memory(>=256GiB))"]
+
+[[benchmark]]
+model = "mlx-community/llama-3.3-70b-instruct-fp16"
+extra_constraints = ["All(Memory(>=256GiB))"]
+
+[[benchmark]]
+model = "mlx-community/GLM-4.5-Air-8bit"
+extra_constraints = ["All(Memory(>=256GiB))"]
+
+[[benchmark]]
+model = "mlx-community/GLM-4.5-Air-bf16"
+extra_constraints = ["All(Memory(>=256GiB))"]
+
+[[benchmark]]
+model = "mlx-community/GLM-4.7-4bit"
+extra_constraints = ["All(Memory(>=256GiB))"]
+
+[[benchmark]]
+model = "mlx-community/MiniMax-M2.1-3bit"
+extra_constraints = ["All(Memory(>=256GiB))"]
+
+[[benchmark]]
+model = "mlx-community/MiniMax-M2.1-8bit"
+extra_constraints = ["All(Memory(>=256GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-235B-A22B-Instruct-2507-4bit"
+extra_constraints = ["All(Memory(>=256GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-Coder-Next-bf16"
+extra_constraints = ["All(Memory(>=256GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Step-3.5-Flash-4bit"
+extra_constraints = ["All(Memory(>=256GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Step-3.5-Flash-6bit"
+extra_constraints = ["All(Memory(>=256GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Step-3.5-Flash-8Bit"
+extra_constraints = ["All(Memory(>=256GiB))"]
+
+[[benchmark]]
+model = "mlx-community/DeepSeek-V3.1-4bit"
+extra_constraints = ["All(Memory(>=512GiB))"]
+
+[[benchmark]]
+model = "mlx-community/GLM-4.7-6bit"
+extra_constraints = ["All(Memory(>=512GiB))"]
+
+[[benchmark]]
+model = "mlx-community/GLM-4.7-8bit-gs32"
+extra_constraints = ["All(Memory(>=512GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-235B-A22B-Instruct-2507-8bit"
+extra_constraints = ["All(Memory(>=512GiB))"]
+
+[[benchmark]]
+model = "mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit"
+extra_constraints = ["All(Memory(>=512GiB))"]
@@ -12,6 +12,8 @@
    ttftMs,
    tps,
    totalTokens,
+    thinkingEnabled as thinkingEnabledStore,
+    setConversationThinking,
  } from "$lib/stores/app.svelte";
  import ChatAttachments from "./ChatAttachments.svelte";
  import ImageParamsPanel from "./ImageParamsPanel.svelte";
@@ -25,6 +27,7 @@
    autofocus?: boolean;
    showModelSelector?: boolean;
    modelTasks?: Record<string, string[]>;
+    modelCapabilities?: Record<string, string[]>;
  }

  let {
@@ -34,6 +37,7 @@
    autofocus = true,
    showModelSelector = false,
    modelTasks = {},
+    modelCapabilities = {},
  }: Props = $props();

  let message = $state("");
@@ -41,6 +45,7 @@
  let fileInputRef: HTMLInputElement | undefined = $state();
  let uploadedFiles = $state<ChatUploadedFile[]>([]);
  let isDragOver = $state(false);
+  const thinkingEnabled = $derived(thinkingEnabledStore());
  let loading = $derived(isLoading());
  const currentModel = $derived(selectedChatModel());
  const instanceData = $derived(instances());
@@ -95,6 +100,12 @@
    );
  });

+  const modelSupportsThinking = $derived(() => {
+    if (!currentModel) return false;
+    const caps = modelCapabilities[currentModel] || [];
+    return caps.includes("thinking") && caps.includes("text");
+  });
+
  const isEditOnlyWithoutImage = $derived(
    currentModel !== null &&
      modelSupportsOnlyImageEditing(currentModel) &&
@@ -254,6 +265,7 @@

  function handleSubmit() {
    if ((!message.trim() && uploadedFiles.length === 0) || loading) return;
+    if (isEditOnlyWithoutImage) return;

    const content = message.trim();
    const files = [...uploadedFiles];
@@ -278,11 +290,19 @@
      if (imageFile.preview) {
        editImage(content, imageFile.preview);
      }
-    } else if (isImageModel() && content) {
+    } else if (
+      currentModel &&
+      modelSupportsTextToImage(currentModel) &&
+      content
+    ) {
      // Use image generation for text-to-image models
      generateImage(content);
    } else {
-      sendMessage(content, files);
+      sendMessage(
+        content,
+        files,
+        modelSupportsThinking() ? thinkingEnabled : null,
+      );
    }

    // Refocus the textarea after sending
@@ -520,6 +540,35 @@
            </div>
          {/if}
        </div>
+        <!-- Thinking toggle -->
+        {#if modelSupportsThinking()}
+          <button
+            type="button"
+            onclick={() => setConversationThinking(!thinkingEnabled)}
+            class="flex items-center gap-1.5 px-2 py-1 rounded text-xs font-mono tracking-wide transition-all duration-200 flex-shrink-0 cursor-pointer border {thinkingEnabled
+              ? 'bg-exo-yellow/15 border-exo-yellow/40 text-exo-yellow'
+              : 'bg-exo-medium-gray/30 border-exo-medium-gray/50 text-exo-light-gray/60 hover:text-exo-light-gray'}"
+            title={thinkingEnabled
+              ? "Thinking enabled — click to disable"
+              : "Thinking disabled — click to enable"}
+          >
+            <svg
+              class="w-3.5 h-3.5"
+              viewBox="0 0 24 24"
+              fill="none"
+              stroke="currentColor"
+              stroke-width="1.5"
+            >
+              <path
+                d="M12 2a7 7 0 0 0-7 7c0 2.38 1.19 4.47 3 5.74V17a1 1 0 0 0 1 1h6a1 1 0 0 0 1-1v-2.26c1.81-1.27 3-3.36 3-5.74a7 7 0 0 0-7-7zM9 20h6M10 22h4"
+                stroke-linecap="round"
+                stroke-linejoin="round"
+              />
+            </svg>
+            <span>{thinkingEnabled ? "THINK" : "NO THINK"}</span>
+          </button>
+        {/if}
+
        <!-- Performance stats -->
        {#if currentTtft !== null || currentTps !== null}
          <div class="flex items-center gap-4 text-xs font-mono flex-shrink-0">
@@ -599,9 +648,8 @@
            : isImageModel()
              ? "Describe the image you want to generate..."
              : placeholder}
-        disabled={loading}
        rows={1}
-        class="flex-1 resize-none bg-transparent text-foreground placeholder:text-exo-light-gray/60 placeholder:text-sm placeholder:tracking-[0.15em] placeholder:leading-7 focus:outline-none focus:ring-0 focus:border-none disabled:opacity-50 text-sm leading-7 font-mono"
+        class="flex-1 resize-none bg-transparent text-foreground placeholder:text-exo-light-gray/60 placeholder:text-sm placeholder:tracking-[0.15em] placeholder:leading-7 focus:outline-none focus:ring-0 focus:border-none text-sm leading-7 font-mono"
        style="min-height: 28px; max-height: 150px;"
      ></textarea>

@@ -225,6 +225,7 @@
  }

  function handleDeleteClick(messageId: string) {
+    if (loading) return;
    deleteConfirmId = messageId;
  }

@@ -255,7 +256,7 @@
 </script>

 <div class="flex flex-col gap-4 sm:gap-6 {className}">
-  {#each messageList as message (message.id)}
+  {#each messageList as message, i (message.id)}
    <div
      class="group flex {message.role === 'user'
        ? 'justify-end'
@@ -317,9 +318,11 @@
          <!-- Delete confirmation -->
          <div class="bg-red-500/10 border border-red-500/30 rounded-lg p-3">
            <p class="text-xs text-red-400 mb-3">
-              Delete this message{message.role === "user"
-                ? " and all responses after it"
-                : ""}?
+              {#if i === messageList.length - 1}
+                Delete this message?
+              {:else}
+                Delete this message and all messages after it?
+              {/if}
            </p>
            <div class="flex gap-2 justify-end">
              <button
@@ -751,8 +754,13 @@
            <!-- Delete button -->
            <button
              onclick={() => handleDeleteClick(message.id)}
-              class="p-1.5 text-exo-light-gray hover:text-red-400 transition-colors rounded hover:bg-red-500/10 cursor-pointer"
-              title="Delete message"
+              disabled={loading}
+              class="p-1.5 transition-colors rounded {loading
+                ? 'text-exo-light-gray/30 cursor-not-allowed'
+                : 'text-exo-light-gray hover:text-red-400 hover:bg-red-500/10 cursor-pointer'}"
+              title={loading
+                ? "Cannot delete while generating"
+                : "Delete message"}
            >
              <svg
                class="w-3.5 h-3.5"
@@ -13,6 +13,12 @@
      d="M12 2l3.09 6.26L22 9.27l-5 4.87 1.18 6.88L12 17.77l-6.18 3.25L7 14.14 2 9.27l6.91-1.01L12 2z"
    />
  </svg>
+{:else if family === "recents"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M13 3a9 9 0 0 0-9 9H1l3.89 3.89.07.14L9 12H6c0-3.87 3.13-7 7-7s7 3.13 7 7-3.13 7-7 7c-1.93 0-3.68-.79-4.94-2.06l-1.42 1.42A8.954 8.954 0 0 0 13 21a9 9 0 0 0 0-18zm-1 5v5l4.28 2.54.72-1.21-3.5-2.08V8H12z"
+    />
+  </svg>
 {:else if family === "llama" || family === "meta"}
  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
    <path
@@ -5,15 +5,22 @@
    families: string[];
    selectedFamily: string | null;
    hasFavorites: boolean;
+    hasRecents: boolean;
    onSelect: (family: string | null) => void;
  };

-  let { families, selectedFamily, hasFavorites, onSelect }: FamilySidebarProps =
-    $props();
+  let {
+    families,
+    selectedFamily,
+    hasFavorites,
+    hasRecents,
+    onSelect,
+  }: FamilySidebarProps = $props();

  // Family display names
  const familyNames: Record<string, string> = {
    favorites: "Favorites",
+    recents: "Recent",
    huggingface: "Hub",
    llama: "Meta",
    qwen: "Qwen",
@@ -89,6 +96,31 @@
    </button>
  {/if}

+  <!-- Recent (only show if has recent models) -->
+  {#if hasRecents}
+    <button
+      type="button"
+      onclick={() => onSelect("recents")}
+      class="group flex flex-col items-center justify-center p-2 rounded transition-all duration-200 cursor-pointer {selectedFamily ===
+      'recents'
+        ? 'bg-exo-yellow/20 border-l-2 border-exo-yellow'
+        : 'hover:bg-white/5 border-l-2 border-transparent'}"
+      title="Recently launched models"
+    >
+      <FamilyLogos
+        family="recents"
+        class={selectedFamily === "recents"
+          ? "text-exo-yellow"
+          : "text-white/50 group-hover:text-white/70"}
+      />
+      <span
+        class="text-[9px] font-mono mt-0.5 {selectedFamily === 'recents'
+          ? 'text-exo-yellow'
+          : 'text-white/40 group-hover:text-white/60'}">Recent</span
+      >
+    </button>
+  {/if}
+
  <!-- HuggingFace Hub -->
  <button
    type="button"
@@ -186,7 +186,9 @@
  // topology payload is missing them. Topology order is preserved exactly so
  // that the mini preview matches the main TopologyGraph layout.
  const nodeList = $derived(() => {
-    const nodesFromTopology = Object.keys(nodes).map((id) => {
+    const nodesFromTopology = Object.keys(nodes)
+      .filter((id) => nodes[id].node_type !== "lite")
+      .map((id) => {
      const info = nodes[id];
      const totalBytes =
        info.macmon_info?.memory?.ram_total ?? info.system_info?.memory ?? 0;
@@ -422,9 +424,16 @@
    const bToACandidates: Array<{ ip: string; iface: string | null }> = [];

    for (const edge of topology.edges) {
-      const ip = edge.sendBackIp || "?";
-      const iface =
-        edge.sendBackInterface || getInterfaceForIp(edge.source, ip);
+      let ip: string;
+      let iface: string | null;
+
+      if (edge.sourceRdmaIface || edge.sinkRdmaIface) {
+        ip = "RDMA";
+        iface = `${edge.sourceRdmaIface || "?"} \u2192 ${edge.sinkRdmaIface || "?"}`;
+      } else {
+        ip = edge.sendBackIp || "?";
+        iface = edge.sendBackInterface || getInterfaceForIp(edge.source, ip);
+      }

      if (edge.source === nodeId1 && edge.target === nodeId2) {
        aToBCandidates.push({ ip, iface });
@@ -26,6 +26,7 @@
    nodeNames: string[];
    nodeIds: string[];
  };
+  type ModelFitStatus = "fits_now" | "fits_cluster_capacity" | "too_large";

  type ModelPickerGroupProps = {
    group: ModelGroup;
@@ -33,11 +34,13 @@
    isFavorite: boolean;
    selectedModelId: string | null;
    canModelFit: (id: string) => boolean;
+    getModelFitStatus: (id: string) => ModelFitStatus;
    onToggleExpand: () => void;
    onSelectModel: (modelId: string) => void;
    onToggleFavorite: (baseModelId: string) => void;
    onShowInfo: (group: ModelGroup) => void;
    downloadStatusMap?: Map<string, DownloadAvailability>;
+    launchedAt?: number;
  };

  let {
@@ -46,11 +49,13 @@
    isFavorite,
    selectedModelId,
    canModelFit,
+    getModelFitStatus,
    onToggleExpand,
    onSelectModel,
    onToggleFavorite,
    onShowInfo,
    downloadStatusMap,
+    launchedAt,
  }: ModelPickerGroupProps = $props();

  // Group-level download status: show if any variant is downloaded
@@ -72,10 +77,45 @@
    return `${mb}MB`;
  }

+  function timeAgo(ts: number): string {
+    const seconds = Math.floor((Date.now() - ts) / 1000);
+    if (seconds < 60) return "just now";
+    const minutes = Math.floor(seconds / 60);
+    if (minutes < 60) return `${minutes}m ago`;
+    const hours = Math.floor(minutes / 60);
+    if (hours < 24) return `${hours}h ago`;
+    const days = Math.floor(hours / 24);
+    return `${days}d ago`;
+  }
+
  // Check if any variant can fit
  const anyVariantFits = $derived(
    group.variants.some((v) => canModelFit(v.id)),
  );
+  const groupFitStatus = $derived.by((): ModelFitStatus => {
+    let hasClusterCapacityOnly = false;
+    for (const variant of group.variants) {
+      const fitStatus = getModelFitStatus(variant.id);
+      if (fitStatus === "fits_now") {
+        return "fits_now";
+      }
+      if (fitStatus === "fits_cluster_capacity") {
+        hasClusterCapacityOnly = true;
+      }
+    }
+    return hasClusterCapacityOnly ? "fits_cluster_capacity" : "too_large";
+  });
+
+  function getSizeClassForFitStatus(fitStatus: ModelFitStatus): string {
+    switch (fitStatus) {
+      case "fits_now":
+        return "text-white/40";
+      case "fits_cluster_capacity":
+        return "text-orange-400/80";
+      case "too_large":
+        return "text-red-400/70";
+    }
+  }

  // Check if this group's model is currently selected (for single-variant groups)
  const isMainSelected = $derived(
@@ -244,7 +284,14 @@

    <!-- Size indicator (smallest variant) -->
    {#if !group.hasMultipleVariants && group.smallestVariant?.storage_size_megabytes}
-      <span class="text-xs font-mono text-white/30 flex-shrink-0">
+      {@const singleVariantFitStatus = getModelFitStatus(
+        group.smallestVariant.id,
+      )}
+      <span
+        class="text-xs font-mono flex-shrink-0 {getSizeClassForFitStatus(
+          singleVariantFitStatus,
+        )}"
+      >
        {formatSize(group.smallestVariant.storage_size_megabytes)}
      </span>
    {/if}
@@ -255,13 +302,24 @@
        .map((v) => v.storage_size_megabytes || 0)
        .filter((s) => s > 0)
        .sort((a, b) => a - b)}
-      <span class="text-xs font-mono text-white/30 flex-shrink-0">
+      <span
+        class="text-xs font-mono flex-shrink-0 {getSizeClassForFitStatus(
+          groupFitStatus,
+        )}"
+      >
        {group.variants.length} variants{#if sizes.length >= 2}{" "}({formatSize(
            sizes[0],
          )}-{formatSize(sizes[sizes.length - 1])}){/if}
      </span>
    {/if}

+    <!-- Time ago (for recent models) -->
+    {#if launchedAt}
+      <span class="text-xs font-mono text-white/20 flex-shrink-0">
+        {timeAgo(launchedAt)}
+      </span>
+    {/if}
+
    <!-- Download availability indicator -->
    {#if groupDownloadStatus && groupDownloadStatus.nodeIds.length > 0}
      <span
@@ -360,6 +418,7 @@
  {#if isExpanded && group.hasMultipleVariants}
    <div class="bg-black/20 border-t border-white/5">
      {#each group.variants as variant}
+        {@const fitStatus = getModelFitStatus(variant.id)}
        {@const modelCanFit = canModelFit(variant.id)}
        {@const isSelected = selectedModelId === variant.id}
        <button
@@ -384,7 +443,11 @@
          </span>

          <!-- Size -->
-          <span class="text-xs font-mono text-white/40 flex-1">
+          <span
+            class="text-xs font-mono flex-1 {getSizeClassForFitStatus(
+              fitStatus,
+            )}"
+          >
            {formatSize(variant.storage_size_megabytes)}
          </span>

@@ -6,6 +6,7 @@
  import ModelFilterPopover from "./ModelFilterPopover.svelte";
  import HuggingFaceResultItem from "./HuggingFaceResultItem.svelte";
  import { getNodesWithModelDownloaded } from "$lib/utils/downloads";
+  import { getRecentEntries } from "$lib/stores/recents.svelte";

  interface ModelInfo {
    id: string;
@@ -46,13 +47,18 @@
    tags: string[];
  }

+  type ModelFitStatus = "fits_now" | "fits_cluster_capacity" | "too_large";
+
  type ModelPickerModalProps = {
    isOpen: boolean;
    models: ModelInfo[];
    selectedModelId: string | null;
    favorites: Set<string>;
+    recentModelIds?: string[];
+    hasRecents?: boolean;
    existingModelIds: Set<string>;
    canModelFit: (modelId: string) => boolean;
+    getModelFitStatus: (modelId: string) => ModelFitStatus;
    onSelect: (modelId: string) => void;
    onClose: () => void;
    onToggleFavorite: (baseModelId: string) => void;
@@ -76,8 +82,11 @@
    models,
    selectedModelId,
    favorites,
+    recentModelIds = [],
+    hasRecents: hasRecentsTab = false,
    existingModelIds,
    canModelFit,
+    getModelFitStatus,
    onSelect,
    onClose,
    onToggleFavorite,
@@ -383,7 +392,11 @@
    // Filter by family
    if (selectedFamily === "favorites") {
      result = result.filter((g) => favorites.has(g.id));
-    } else if (selectedFamily && selectedFamily !== "huggingface") {
+    } else if (
+      selectedFamily &&
+      selectedFamily !== "huggingface" &&
+      selectedFamily !== "recents"
+    ) {
      result = result.filter((g) => g.family === selectedFamily);
    }

@@ -427,13 +440,23 @@
      );
    }

-    // Sort: models that fit first, then by size (largest first)
+    // Sort: fits-now first, then fits-cluster-capacity, then too-large
    result.sort((a, b) => {
-      const aFits = a.variants.some((v) => canModelFit(v.id));
-      const bFits = b.variants.some((v) => canModelFit(v.id));
+      const getGroupFitRank = (group: ModelGroup): number => {
+        let hasClusterCapacityOnly = false;
+        for (const variant of group.variants) {
+          const fitStatus = getModelFitStatus(variant.id);
+          if (fitStatus === "fits_now") return 0;
+          if (fitStatus === "fits_cluster_capacity") {
+            hasClusterCapacityOnly = true;
+          }
+        }
+        return hasClusterCapacityOnly ? 1 : 2;
+      };

-      if (aFits && !bFits) return -1;
-      if (!aFits && bFits) return 1;
+      const aRank = getGroupFitRank(a);
+      const bRank = getGroupFitRank(b);
+      if (aRank !== bRank) return aRank - bRank;

      return (
        (b.smallestVariant.storage_size_megabytes || 0) -
@@ -447,6 +470,48 @@
  // Check if any favorites exist
  const hasFavorites = $derived(favorites.size > 0);

+  // Timestamp lookup for recent models
+  const recentTimestamps = $derived(
+    new Map(getRecentEntries().map((e) => [e.modelId, e.launchedAt])),
+  );
+
+  // Recent models: single-variant ModelGroups in launch order
+  const recentGroups = $derived.by((): ModelGroup[] => {
+    if (!recentModelIds || recentModelIds.length === 0) return [];
+    const result: ModelGroup[] = [];
+    for (const id of recentModelIds) {
+      const model = models.find((m) => m.id === id);
+      if (model) {
+        result.push({
+          id: model.base_model || model.id,
+          name: model.name || model.id,
+          capabilities: model.capabilities || ["text"],
+          family: model.family || "",
+          variants: [model],
+          smallestVariant: model,
+          hasMultipleVariants: false,
+        });
+      }
+    }
+    return result;
+  });
+
+  // Filtered recent groups (apply search query)
+  const filteredRecentGroups = $derived.by((): ModelGroup[] => {
+    if (!searchQuery.trim()) return recentGroups;
+    const query = searchQuery.toLowerCase().trim();
+    return recentGroups.filter(
+      (g) =>
+        g.name.toLowerCase().includes(query) ||
+        g.variants.some(
+          (v) =>
+            v.id.toLowerCase().includes(query) ||
+            (v.name || "").toLowerCase().includes(query) ||
+            (v.quantization || "").toLowerCase().includes(query),
+        ),
+    );
+  });
+
  function toggleGroupExpanded(groupId: string) {
    const next = new Set(expandedGroups);
    if (next.has(groupId)) {
@@ -604,6 +669,7 @@
        families={uniqueFamilies}
        {selectedFamily}
        {hasFavorites}
+        hasRecents={hasRecentsTab}
        onSelect={(family) => (selectedFamily = family)}
      />

@@ -711,6 +777,45 @@
              </div>
            </div>
          </div>
+        {:else if selectedFamily === "recents"}
+          <!-- Recent models view -->
+          {#if filteredRecentGroups.length === 0}
+            <div
+              class="flex flex-col items-center justify-center h-full text-white/40 p-8"
+            >
+              <svg
+                class="w-12 h-12 mb-3"
+                viewBox="0 0 24 24"
+                fill="currentColor"
+              >
+                <path
+                  d="M13 3a9 9 0 0 0-9 9H1l3.89 3.89.07.14L9 12H6c0-3.87 3.13-7 7-7s7 3.13 7 7-3.13 7-7 7c-1.93 0-3.68-.79-4.94-2.06l-1.42 1.42A8.954 8.954 0 0 0 13 21a9 9 0 0 0 0-18zm-1 5v5l4.28 2.54.72-1.21-3.5-2.08V8H12z"
+                />
+              </svg>
+              <p class="font-mono text-sm">
+                {searchQuery
+                  ? "No matching recent models"
+                  : "No recently launched models"}
+              </p>
+            </div>
+          {:else}
+            {#each filteredRecentGroups as group}
+              <ModelPickerGroup
+                {group}
+                isExpanded={expandedGroups.has(group.id)}
+                isFavorite={favorites.has(group.id)}
+                {selectedModelId}
+                {canModelFit}
+                {getModelFitStatus}
+                onToggleExpand={() => toggleGroupExpanded(group.id)}
+                onSelectModel={handleSelect}
+                {onToggleFavorite}
+                onShowInfo={(g) => (infoGroup = g)}
+                downloadStatusMap={getVariantDownloadMap(group)}
+                launchedAt={recentTimestamps.get(group.variants[0]?.id ?? "")}
+              />
+            {/each}
+          {/if}
        {:else if filteredGroups.length === 0}
          <div
            class="flex flex-col items-center justify-center h-full text-white/40 p-8"
@@ -742,6 +847,7 @@
              isFavorite={favorites.has(group.id)}
              {selectedModelId}
              {canModelFit}
+              {getModelFitStatus}
              onToggleExpand={() => toggleGroupExpanded(group.id)}
              onSelectModel={handleSelect}
              {onToggleFavorite}
@@ -6,6 +6,8 @@
    isTopologyMinimized,
    debugMode,
    nodeThunderboltBridge,
+    nodeRdmaCtl,
+    nodeIdentities,
    type NodeInfo,
  } from "$lib/stores/app.svelte";

@@ -31,6 +33,8 @@
  const data = $derived(topologyData());
  const debugEnabled = $derived(debugMode());
  const tbBridgeData = $derived(nodeThunderboltBridge());
+  const rdmaCtlData = $derived(nodeRdmaCtl());
+  const identitiesData = $derived(nodeIdentities());

  function getNodeLabel(nodeId: string): string {
    const node = data?.nodes?.[nodeId];
@@ -333,14 +337,27 @@
      if (edge.source === a) entry.aToB = true;
      else entry.bToA = true;

-      const ip = edge.sendBackIp || "?";
-      const ifaceInfo = getInterfaceLabel(edge.source, ip);
+      let ip: string;
+      let ifaceLabel: string;
+      let missingIface: boolean;
+
+      if (edge.sourceRdmaIface || edge.sinkRdmaIface) {
+        ip = "RDMA";
+        ifaceLabel = `${edge.sourceRdmaIface || "?"} \u2192 ${edge.sinkRdmaIface || "?"}`;
+        missingIface = false;
+      } else {
+        ip = edge.sendBackIp || "?";
+        const ifaceInfo = getInterfaceLabel(edge.source, ip);
+        ifaceLabel = ifaceInfo.label;
+        missingIface = ifaceInfo.missing;
+      }
+
      entry.connections.push({
        from: edge.source,
        to: edge.target,
        ip,
-        ifaceLabel: ifaceInfo.label,
-        missingIface: ifaceInfo.missing,
+        ifaceLabel,
+        missingIface,
      });
      pairMap.set(key, entry);
    });
@@ -537,6 +554,7 @@
      const clipPathId = `clip-${nodeInfo.id.replace(/[^a-zA-Z0-9]/g, "-")}`;

      const modelLower = modelId.toLowerCase();
+      const isLiteNode = node.node_type === "lite";

      // Check node states for styling
      const isHighlighted = highlightedNodes.has(nodeInfo.id);
@@ -889,6 +907,94 @@
          .attr("height", trackpadHeight)
          .attr("fill", "rgba(255,255,255,0.08)")
          .attr("rx", 2);
+      } else if (modelLower.includes("iphone")) {
+        // iPhone - rounded rectangle phone shape with Dynamic Island
+        iconBaseWidth = nodeRadius * 0.7;
+        iconBaseHeight = nodeRadius * 1.4;
+        const x = nodeInfo.x - iconBaseWidth / 2;
+        const y = nodeInfo.y - iconBaseHeight / 2;
+        const cornerRadius = iconBaseWidth * 0.2;
+        const bezel = 3;
+
+        // Phone body (outer frame)
+        nodeG
+          .append("rect")
+          .attr("class", "node-outline")
+          .attr("x", x)
+          .attr("y", y)
+          .attr("width", iconBaseWidth)
+          .attr("height", iconBaseHeight)
+          .attr("rx", cornerRadius)
+          .attr("fill", "#1a1a1a")
+          .attr("stroke", wireColor)
+          .attr("stroke-width", strokeWidth);
+
+        // Screen area (inner)
+        const screenClipId = `iphone-clip-${nodeInfo.id.replace(/[^a-zA-Z0-9]/g, "-")}`;
+        defs
+          .append("clipPath")
+          .attr("id", screenClipId)
+          .append("rect")
+          .attr("x", x + bezel)
+          .attr("y", y + bezel)
+          .attr("width", iconBaseWidth - bezel * 2)
+          .attr("height", iconBaseHeight - bezel * 2)
+          .attr("rx", cornerRadius - 1);
+
+        nodeG
+          .append("rect")
+          .attr("x", x + bezel)
+          .attr("y", y + bezel)
+          .attr("width", iconBaseWidth - bezel * 2)
+          .attr("height", iconBaseHeight - bezel * 2)
+          .attr("rx", cornerRadius - 1)
+          .attr("fill", screenFill);
+
+        // Memory fill on screen (fills from bottom up)
+        if (ramUsagePercent > 0) {
+          const memFillTotalHeight = iconBaseHeight - bezel * 2;
+          const memFillActualHeight =
+            (ramUsagePercent / 100) * memFillTotalHeight;
+          nodeG
+            .append("rect")
+            .attr("x", x + bezel)
+            .attr(
+              "y",
+              y + bezel + (memFillTotalHeight - memFillActualHeight),
+            )
+            .attr("width", iconBaseWidth - bezel * 2)
+            .attr("height", memFillActualHeight)
+            .attr("fill", "rgba(255,215,0,0.85)")
+            .attr("clip-path", `url(#${screenClipId})`);
+        }
+
+        // Dynamic Island notch (centered near top)
+        const diWidth = iconBaseWidth * 0.3;
+        const diHeight = iconBaseHeight * 0.04;
+        nodeG
+          .append("rect")
+          .attr("x", nodeInfo.x - diWidth / 2)
+          .attr("y", y + bezel + 4)
+          .attr("width", diWidth)
+          .attr("height", diHeight)
+          .attr("rx", diHeight / 2)
+          .attr("fill", "#000000");
+
+        // Apple logo on screen (centered)
+        const targetLogoHeight = iconBaseHeight * 0.14;
+        const logoScale = targetLogoHeight / LOGO_NATIVE_HEIGHT;
+        const logoX = nodeInfo.x - (LOGO_NATIVE_WIDTH * logoScale) / 2;
+        const logoY =
+          nodeInfo.y - (LOGO_NATIVE_HEIGHT * logoScale) / 2;
+        nodeG
+          .append("path")
+          .attr("d", APPLE_LOGO_PATH)
+          .attr(
+            "transform",
+            `translate(${logoX}, ${logoY}) scale(${logoScale})`,
+          )
+          .attr("fill", "#FFFFFF")
+          .attr("opacity", 0.9);
      } else {
        // Default/Unknown - holographic hexagon
        const hexRadius = nodeRadius * 0.6;
@@ -907,9 +1013,43 @@
          .attr("stroke-width", strokeWidth);
      }

+      // --- LITE badge for lite nodes ---
+      if (isLiteNode) {
+        const badgeX = nodeInfo.x + iconBaseWidth / 2 + 4;
+        const badgeY = nodeInfo.y - iconBaseHeight / 2 - 2;
+        const badgeFontSize = Math.max(9, nodeRadius * 0.12);
+        const badgePadH = 4;
+        const badgePadV = 2;
+        const badgeWidth = badgeFontSize * 2.8 + badgePadH * 2;
+        const badgeHeight = badgeFontSize + badgePadV * 2;
+
+        nodeG
+          .append("rect")
+          .attr("x", badgeX)
+          .attr("y", badgeY)
+          .attr("width", badgeWidth)
+          .attr("height", badgeHeight)
+          .attr("rx", 3)
+          .attr("fill", "rgba(255,215,0,0.15)")
+          .attr("stroke", "rgba(255,215,0,0.6)")
+          .attr("stroke-width", 1);
+
+        nodeG
+          .append("text")
+          .attr("x", badgeX + badgeWidth / 2)
+          .attr("y", badgeY + badgeHeight / 2)
+          .attr("text-anchor", "middle")
+          .attr("dominant-baseline", "central")
+          .attr("fill", "rgba(255,215,0,0.9)")
+          .attr("font-size", badgeFontSize)
+          .attr("font-weight", "700")
+          .attr("font-family", "SF Mono, Monaco, monospace")
+          .text("LITE");
+      }
+
      // --- Vertical GPU Bar (right side of icon) ---
-      // Show in both full mode and minimized mode (scaled appropriately)
-      if (showFullLabels || isMinimized) {
+      // Show in both full mode and minimized mode (scaled appropriately), but not for lite nodes
+      if ((showFullLabels || isMinimized) && !isLiteNode) {
        const gpuBarWidth = isMinimized
          ? Math.max(16, nodeRadius * 0.32)
          : Math.max(28, nodeRadius * 0.3);
@@ -1120,15 +1260,17 @@
          .text(` (${ramUsagePercent.toFixed(0)}%)`);
      }

-      // Debug mode: Show TB bridge status
+      // Debug mode: Show TB bridge and RDMA status
      if (debugEnabled) {
+        let debugLabelY =
+          nodeInfo.y +
+          iconBaseHeight / 2 +
+          (showFullLabels ? 32 : showCompactLabels ? 26 : 22);
+        const debugFontSize = showFullLabels ? 9 : 7;
+        const debugLineHeight = showFullLabels ? 11 : 9;
+
        const tbStatus = tbBridgeData[nodeInfo.id];
        if (tbStatus) {
-          const tbY =
-            nodeInfo.y +
-            iconBaseHeight / 2 +
-            (showFullLabels ? 32 : showCompactLabels ? 26 : 22);
-          const tbFontSize = showFullLabels ? 9 : 7;
          const tbColor = tbStatus.enabled
            ? "rgba(234,179,8,0.9)"
            : "rgba(100,100,100,0.7)";
@@ -1136,12 +1278,46 @@
          nodeG
            .append("text")
            .attr("x", nodeInfo.x)
-            .attr("y", tbY)
+            .attr("y", debugLabelY)
            .attr("text-anchor", "middle")
            .attr("fill", tbColor)
-            .attr("font-size", tbFontSize)
+            .attr("font-size", debugFontSize)
            .attr("font-family", "SF Mono, Monaco, monospace")
            .text(tbText);
+          debugLabelY += debugLineHeight;
+        }
+
+        const rdmaStatus = rdmaCtlData[nodeInfo.id];
+        if (rdmaStatus !== undefined) {
+          const rdmaColor = rdmaStatus.enabled
+            ? "rgba(74,222,128,0.9)"
+            : "rgba(100,100,100,0.7)";
+          const rdmaText = rdmaStatus.enabled ? "RDMA:ON" : "RDMA:OFF";
+          nodeG
+            .append("text")
+            .attr("x", nodeInfo.x)
+            .attr("y", debugLabelY)
+            .attr("text-anchor", "middle")
+            .attr("fill", rdmaColor)
+            .attr("font-size", debugFontSize)
+            .attr("font-family", "SF Mono, Monaco, monospace")
+            .text(rdmaText);
+          debugLabelY += debugLineHeight;
+        }
+
+        const identity = identitiesData[nodeInfo.id];
+        if (identity?.osVersion) {
+          nodeG
+            .append("text")
+            .attr("x", nodeInfo.x)
+            .attr("y", debugLabelY)
+            .attr("text-anchor", "middle")
+            .attr("fill", "rgba(179,179,179,0.7)")
+            .attr("font-size", debugFontSize)
+            .attr("font-family", "SF Mono, Monaco, monospace")
+            .text(
+              `macOS ${identity.osVersion}${identity.osBuildVersion ? ` (${identity.osBuildVersion})` : ""}`,
+            );
        }
      }
    });
@@ -49,6 +49,8 @@ export interface NodeInfo {
  };
  last_macmon_update: number;
  friendly_name?: string;
+  os_version?: string;
+  node_type?: string;
 }

 export interface TopologyEdge {
@@ -56,6 +58,8 @@ export interface TopologyEdge {
  target: string;
  sendBackIp?: string;
  sendBackInterface?: string;
+  sourceRdmaIface?: string;
+  sinkRdmaIface?: string;
 }

 export interface TopologyData {
@@ -76,6 +80,9 @@ interface RawNodeIdentity {
  modelId?: string;
  chipId?: string;
  friendlyName?: string;
+  osVersion?: string;
+  osBuildVersion?: string;
+  nodeType?: string;
 }

 interface RawMemoryUsage {
@@ -225,6 +232,19 @@ interface RawStateResponse {
  nodeMemory?: Record<string, RawMemoryUsage>;
  nodeSystem?: Record<string, RawSystemPerformanceProfile>;
  nodeNetwork?: Record<string, RawNodeNetworkInfo>;
+  // Thunderbolt identifiers per node
+  nodeThunderbolt?: Record<
+    string,
+    {
+      interfaces: Array<{
+        rdmaInterface: string;
+        domainUuid: string;
+        linkSpeed: string;
+      }>;
+    }
+  >;
+  // RDMA ctl status per node
+  nodeRdmaCtl?: Record<string, { enabled: boolean }>;
  // Thunderbolt bridge status per node
  nodeThunderboltBridge?: Record<
    string,
@@ -278,6 +298,7 @@ export interface Conversation {
  modelId: string | null;
  sharding: string | null;
  instanceType: string | null;
+  enableThinking: boolean | null;
 }

 const STORAGE_KEY = "exo-conversations";
@@ -425,6 +446,8 @@ function transformTopology(
      },
      last_macmon_update: Date.now() / 1000,
      friendly_name: identity?.friendlyName,
+      os_version: identity?.osVersion,
+      node_type: identity?.nodeType,
    };
  }

@@ -437,6 +460,8 @@ function transformTopology(
        if (!Array.isArray(edgeList)) continue;
        for (const edge of edgeList) {
          let sendBackIp: string | undefined;
+          let sourceRdmaIface: string | undefined;
+          let sinkRdmaIface: string | undefined;
          if (edge && typeof edge === "object" && "sinkMultiaddr" in edge) {
            const multiaddr = edge.sinkMultiaddr;
            if (multiaddr) {
@@ -444,10 +469,23 @@ function transformTopology(
                multiaddr.ip_address ||
                extractIpFromMultiaddr(multiaddr.address);
            }
+          } else if (
+            edge &&
+            typeof edge === "object" &&
+            "sourceRdmaIface" in edge
+          ) {
+            sourceRdmaIface = edge.sourceRdmaIface;
+            sinkRdmaIface = edge.sinkRdmaIface;
          }

          if (nodes[source] && nodes[sink] && source !== sink) {
-            edges.push({ source, target: sink, sendBackIp });
+            edges.push({
+              source,
+              target: sink,
+              sendBackIp,
+              sourceRdmaIface,
+              sinkRdmaIface,
+            });
          }
        }
      }
@@ -490,12 +528,32 @@ class AppStore {
  instances = $state<Record<string, unknown>>({});
  runners = $state<Record<string, unknown>>({});
  downloads = $state<Record<string, unknown[]>>({});
+  nodeDisk = $state<
+    Record<
+      string,
+      { total: { inBytes: number }; available: { inBytes: number } }
+    >
+  >({});
  placementPreviews = $state<PlacementPreview[]>([]);
  selectedPreviewModelId = $state<string | null>(null);
  isLoadingPreviews = $state(false);
  previewNodeFilter = $state<Set<string>>(new Set());
  lastUpdate = $state<number | null>(null);
+  nodeIdentities = $state<Record<string, RawNodeIdentity>>({});
  thunderboltBridgeCycles = $state<string[][]>([]);
+  nodeThunderbolt = $state<
+    Record<
+      string,
+      {
+        interfaces: Array<{
+          rdmaInterface: string;
+          domainUuid: string;
+          linkSpeed: string;
+        }>;
+      }
+    >
+  >({});
+  nodeRdmaCtl = $state<Record<string, { enabled: boolean }>>({});
  nodeThunderboltBridge = $state<
    Record<
      string,
@@ -551,6 +609,7 @@ class AppStore {
          modelId: conversation.modelId ?? null,
          sharding: conversation.sharding ?? null,
          instanceType: conversation.instanceType ?? null,
+          enableThinking: conversation.enableThinking ?? null,
        }));
      }
    } catch (error) {
@@ -740,6 +799,7 @@ class AppStore {
      modelId: derivedModelId,
      sharding: derivedSharding,
      instanceType: derivedInstanceType,
+      enableThinking: null,
    };

    this.conversations.unshift(conversation);
@@ -765,6 +825,7 @@ class AppStore {
    this.hasStartedChat = true;
    this.isTopologyMinimized = true;
    this.isSidebarOpen = true; // Auto-open sidebar when chatting
+    this.thinkingEnabled = conversation.enableThinking ?? true;
    this.refreshConversationModelFromInstances();

    return true;
@@ -1206,6 +1267,15 @@ class AppStore {
      if (data.downloads) {
        this.downloads = data.downloads;
      }
+      if (data.nodeDisk) {
+        this.nodeDisk = data.nodeDisk;
+      }
+      // Node identities (for OS version mismatch detection)
+      this.nodeIdentities = data.nodeIdentities ?? {};
+      // Thunderbolt identifiers per node
+      this.nodeThunderbolt = data.nodeThunderbolt ?? {};
+      // RDMA ctl status per node
+      this.nodeRdmaCtl = data.nodeRdmaCtl ?? {};
      // Thunderbolt bridge cycles
      this.thunderboltBridgeCycles = data.thunderboltBridgeCycles ?? [];
      // Thunderbolt bridge status per node
@@ -1869,6 +1939,11 @@ class AppStore {
    }
  }

+  /**
+   * Whether thinking is enabled for the current conversation
+   */
+  thinkingEnabled = $state(true);
+
  /**
   * Selected model for chat (can be set by the UI)
   */
@@ -2047,6 +2122,7 @@ class AppStore {
      textContent?: string;
      preview?: string;
    }[],
+    enableThinking?: boolean | null,
  ): Promise<void> {
    if ((!content.trim() && (!files || files.length === 0)) || this.isLoading)
      return;
@@ -2194,6 +2270,9 @@ class AppStore {
          stream: true,
          logprobs: true,
          top_logprobs: 5,
+          ...(enableThinking != null && {
+            enable_thinking: enableThinking,
+          }),
        }),
      });

@@ -2852,6 +2931,18 @@ class AppStore {
    );
  }

+  /**
+   * Update the thinking preference for the active conversation
+   */
+  setConversationThinking(enabled: boolean) {
+    this.thinkingEnabled = enabled;
+    const conv = this.getActiveConversation();
+    if (conv) {
+      conv.enableThinking = enabled;
+      this.saveConversationsToStorage();
+    }
+  }
+
  /**
   * Start a download on a specific node
   */
@@ -2958,12 +3049,14 @@ export const topologyData = () => appStore.topologyData;
 export const instances = () => appStore.instances;
 export const runners = () => appStore.runners;
 export const downloads = () => appStore.downloads;
+export const nodeDisk = () => appStore.nodeDisk;
 export const placementPreviews = () => appStore.placementPreviews;
 export const selectedPreviewModelId = () => appStore.selectedPreviewModelId;
 export const isLoadingPreviews = () => appStore.isLoadingPreviews;
 export const lastUpdate = () => appStore.lastUpdate;
 export const isTopologyMinimized = () => appStore.isTopologyMinimized;
 export const selectedChatModel = () => appStore.selectedChatModel;
+export const thinkingEnabled = () => appStore.thinkingEnabled;
 export const debugMode = () => appStore.getDebugMode();
 export const topologyOnlyMode = () => appStore.getTopologyOnlyMode();
 export const chatSidebarVisible = () => appStore.getChatSidebarVisible();
@@ -2979,7 +3072,8 @@ export const sendMessage = (
    textContent?: string;
    preview?: string;
  }[],
-) => appStore.sendMessage(content, files);
+  enableThinking?: boolean | null,
+) => appStore.sendMessage(content, files, enableThinking);
 export const generateImage = (prompt: string, modelId?: string) =>
  appStore.generateImage(prompt, modelId);
 export const editImage = (
@@ -3022,6 +3116,8 @@ export const deleteAllConversations = () => appStore.deleteAllConversations();
 export const renameConversation = (id: string, name: string) =>
  appStore.renameConversation(id, name);
 export const getActiveConversation = () => appStore.getActiveConversation();
+export const setConversationThinking = (enabled: boolean) =>
+  appStore.setConversationThinking(enabled);

 // Sidebar actions
 export const isSidebarOpen = () => appStore.isSidebarOpen;
@@ -3038,7 +3134,12 @@ export const setChatSidebarVisible = (visible: boolean) =>
  appStore.setChatSidebarVisible(visible);
 export const refreshState = () => appStore.fetchState();

-// Thunderbolt bridge status
+// Node identities (for OS version mismatch detection)
+export const nodeIdentities = () => appStore.nodeIdentities;
+
+// Thunderbolt & RDMA status
+export const nodeThunderbolt = () => appStore.nodeThunderbolt;
+export const nodeRdmaCtl = () => appStore.nodeRdmaCtl;
 export const thunderboltBridgeCycles = () => appStore.thunderboltBridgeCycles;
 export const nodeThunderboltBridge = () => appStore.nodeThunderboltBridge;

@@ -0,0 +1,75 @@
+/**
+ * RecentsStore - Manages recently launched models with localStorage persistence
+ */
+
+import { browser } from "$app/environment";
+
+const RECENTS_KEY = "exo-recent-models";
+const MAX_RECENT_MODELS = 20;
+
+interface RecentEntry {
+  modelId: string;
+  launchedAt: number;
+}
+
+class RecentsStore {
+  recents = $state<RecentEntry[]>([]);
+
+  constructor() {
+    if (browser) {
+      this.loadFromStorage();
+    }
+  }
+
+  private loadFromStorage() {
+    try {
+      const stored = localStorage.getItem(RECENTS_KEY);
+      if (stored) {
+        const parsed = JSON.parse(stored) as RecentEntry[];
+        this.recents = parsed;
+      }
+    } catch (error) {
+      console.error("Failed to load recent models:", error);
+    }
+  }
+
+  private saveToStorage() {
+    try {
+      localStorage.setItem(RECENTS_KEY, JSON.stringify(this.recents));
+    } catch (error) {
+      console.error("Failed to save recent models:", error);
+    }
+  }
+
+  recordLaunch(modelId: string) {
+    // Remove existing entry for this model (if any) to move it to top
+    const filtered = this.recents.filter((r) => r.modelId !== modelId);
+    // Prepend new entry
+    const next = [{ modelId, launchedAt: Date.now() }, ...filtered];
+    // Cap at max
+    this.recents = next.slice(0, MAX_RECENT_MODELS);
+    this.saveToStorage();
+  }
+
+  getRecentModelIds(): string[] {
+    return this.recents.map((r) => r.modelId);
+  }
+
+  hasAny(): boolean {
+    return this.recents.length > 0;
+  }
+
+  clearAll() {
+    this.recents = [];
+    this.saveToStorage();
+  }
+}
+
+export const recentsStore = new RecentsStore();
+
+export const hasRecents = () => recentsStore.hasAny();
+export const getRecentModelIds = () => recentsStore.getRecentModelIds();
+export const getRecentEntries = () => recentsStore.recents;
+export const recordRecentLaunch = (modelId: string) =>
+  recentsStore.recordLaunch(modelId);
+export const clearRecents = () => recentsStore.clearAll();
@@ -12,6 +12,11 @@
    toggleFavorite,
    getFavoritesSet,
  } from "$lib/stores/favorites.svelte";
+  import {
+    hasRecents,
+    getRecentModelIds,
+    recordRecentLaunch,
+  } from "$lib/stores/recents.svelte";
  import {
    hasStartedChat,
    isTopologyMinimized,
@@ -37,8 +42,11 @@
    toggleTopologyOnlyMode,
    chatSidebarVisible,
    toggleChatSidebarVisible,
+    nodeThunderbolt,
+    nodeRdmaCtl,
    thunderboltBridgeCycles,
    nodeThunderboltBridge,
+    nodeIdentities,
    type DownloadProgress,
    type PlacementPreview,
  } from "$lib/stores/app.svelte";
@@ -62,8 +70,51 @@
  const sidebarVisible = $derived(chatSidebarVisible());
  const tbBridgeCycles = $derived(thunderboltBridgeCycles());
  const tbBridgeData = $derived(nodeThunderboltBridge());
+  const identitiesData = $derived(nodeIdentities());
+  const tbIdentifiers = $derived(nodeThunderbolt());
+  const rdmaCtlData = $derived(nodeRdmaCtl());
  const nodeFilter = $derived(previewNodeFilter());

+  // Detect macOS version mismatches across cluster nodes
+  const macosVersionMismatch = $derived.by(() => {
+    if (!identitiesData) return null;
+    const entries = Object.entries(identitiesData);
+    // Filter to full macOS nodes (version starts with a digit, e.g. "15.3"), excluding lite nodes
+    const macosNodes = entries.filter(([_, id]) => {
+      if (id.nodeType === "lite") return false;
+      const v = id.osVersion;
+      return v && v !== "Unknown" && /^\d/.test(v);
+    });
+    if (macosNodes.length < 2) return null;
+    // Compare on buildVersion for precise mismatch detection
+    const buildVersions = new Set(
+      macosNodes.map(([_, id]) => id.osBuildVersion ?? id.osVersion),
+    );
+    if (buildVersions.size <= 1) return null;
+    return macosNodes.map(([nodeId, id]) => ({
+      nodeId,
+      friendlyName: getNodeName(nodeId),
+      version: id.osVersion!,
+      buildVersion: id.osBuildVersion ?? "Unknown",
+    }));
+  });
+
+  // Detect TB5 nodes where RDMA is not enabled
+  const tb5WithoutRdma = $derived.by(() => {
+    const rdmaCtl = rdmaCtlData;
+    if (!rdmaCtl) return false;
+    const ids = tbIdentifiers;
+    if (!ids) return false;
+    // Find nodes with TB5 hardware (any TB interface)
+    const tb5NodeIds = Object.entries(ids)
+      .filter(([_, node]) => node.interfaces.length > 0)
+      .map(([id]) => id);
+    if (tb5NodeIds.length < 2) return false;
+    // At least one TB5 node has RDMA disabled
+    return tb5NodeIds.some((id) => rdmaCtl[id]?.enabled !== true);
+  });
+  let tb5InfoDismissed = $state(false);
+
  // Helper to get friendly node name from node ID
  function getNodeName(nodeId: string): string {
    const node = data?.nodes?.[nodeId];
@@ -96,6 +147,12 @@
    }
  }

+  // Warning icon SVG path (reused across warning snippets)
+  const warningIconPath =
+    "M12 9v2m0 4h.01m-6.938 4h13.856c1.54 0 2.502-1.667 1.732-3L13.732 4c-.77-1.333-2.694-1.333-3.464 0L3.34 16c-.77 1.333.192 3 1.732 3z";
+  const infoIconPath =
+    "M13 16h-1v-4h-1m1-4h.01M21 12a9 9 0 11-18 0 9 9 0 0118 0z";
+
  let mounted = $state(false);

  // Instance launch state
@@ -113,6 +170,10 @@
      capabilities?: string[];
    }>
  >([]);
+  type ModelMemoryFitStatus =
+    | "fits_now"
+    | "fits_cluster_capacity"
+    | "too_large";

  // Model tasks lookup for ChatForm - maps both short IDs and full HuggingFace IDs
  const modelTasks = $derived(() => {
@@ -130,6 +191,19 @@
    return tasks;
  });

+  const modelCapabilities = $derived(() => {
+    const caps: Record<string, string[]> = {};
+    for (const model of models) {
+      if (model.capabilities && model.capabilities.length > 0) {
+        caps[model.id] = model.capabilities;
+        if (model.hugging_face_id) {
+          caps[model.hugging_face_id] = model.capabilities;
+        }
+      }
+    }
+    return caps;
+  });
+
  // Helper to check if a model supports image generation
  function modelSupportsImageGeneration(modelId: string): boolean {
    const model = models.find(
@@ -228,6 +302,10 @@
  // Favorites state (reactive)
  const favoritesSet = $derived(getFavoritesSet());

+  // Recent models state (reactive)
+  const recentModelIds = $derived(getRecentModelIds());
+  const showRecentsTab = $derived(hasRecents());
+
  // Slider dragging state
  let isDraggingSlider = $state(false);
  let sliderTrackElement: HTMLDivElement | null = $state(null);
@@ -450,14 +528,41 @@
    );
  });

+  // Calculate total memory in the cluster (in GB)
+  const clusterTotalMemoryGB = $derived(() => {
+    if (!data) return 0;
+    return (
+      Object.values(data.nodes).reduce((acc, n) => {
+        const total =
+          n.macmon_info?.memory?.ram_total ?? n.system_info?.memory ?? 0;
+        return acc + total;
+      }, 0) /
+      (1024 * 1024 * 1024)
+    );
+  });
+
+  function getModelMemoryFitStatus(model: {
+    id: string;
+    name?: string;
+    storage_size_megabytes?: number;
+  }): ModelMemoryFitStatus {
+    const modelSizeGB = getModelSizeGB(model);
+    if (modelSizeGB <= availableMemoryGB()) {
+      return "fits_now";
+    }
+    if (modelSizeGB <= clusterTotalMemoryGB()) {
+      return "fits_cluster_capacity";
+    }
+    return "too_large";
+  }
+
  // Check if a model has enough memory to run
  function hasEnoughMemory(model: {
    id: string;
    name?: string;
    storage_size_megabytes?: number;
  }): boolean {
-    const modelSizeGB = getModelSizeGB(model);
-    return modelSizeGB <= availableMemoryGB();
+    return getModelMemoryFitStatus(model) === "fits_now";
  }

  // Sorted models for dropdown - biggest first, unrunnable at the end
@@ -534,7 +639,9 @@
        models = data.data || [];
        // Restore last launch defaults if available
        const currentNodeCount = topologyData()
-          ? Object.keys(topologyData()!.nodes).length
+          ? Object.values(topologyData()!.nodes).filter(
+              (n) => n.node_type !== "lite",
+            ).length
          : 1;
        applyLaunchDefaults(models, currentNodeCount);
      }
@@ -630,6 +737,9 @@
        // Always auto-select the newly launched model so the user chats to what they just launched
        setSelectedChatModel(modelId);

+        // Record the launch in recent models history
+        recordRecentLaunch(modelId);
+
        // Scroll to the bottom of instances container to show the new instance
        // Use multiple attempts to ensure DOM has updated with the new instance
        const scrollToBottom = () => {
@@ -1539,7 +1649,12 @@
    saveLaunchDefaults();
  }

-  const nodeCount = $derived(data ? Object.keys(data.nodes).length : 0);
+  const totalNodeCount = $derived(data ? Object.keys(data.nodes).length : 0);
+  const fullNodeCount = $derived(
+    data
+      ? Object.values(data.nodes).filter((n) => n.node_type !== "lite").length
+      : 0,
+  );
  const instanceCount = $derived(Object.keys(instanceData).length);

  // Helper to get the number of nodes in a placement preview
@@ -1552,7 +1667,7 @@
  }

  // Available min nodes options based on topology (like old dashboard)
-  const availableMinNodes = $derived(Math.max(1, nodeCount));
+  const availableMinNodes = $derived(Math.max(1, fullNodeCount));

  // Compute which min node values have valid previews for the current model/sharding/instance type
  // A minNodes value N is valid if there exists a placement with nodeCount >= N
@@ -1645,18 +1760,263 @@
  // Calculate total memory usage across all nodes
  const clusterMemory = $derived(() => {
    if (!data) return { used: 0, total: 0 };
-    return Object.values(data.nodes).reduce(
-      (acc, n) => {
-        const total =
-          n.macmon_info?.memory?.ram_total ?? n.system_info?.memory ?? 0;
-        const used = n.macmon_info?.memory?.ram_usage ?? 0;
-        return { used: acc.used + used, total: acc.total + total };
-      },
-      { used: 0, total: 0 },
-    );
+    return Object.values(data.nodes)
+      .filter((n) => n.node_type !== "lite")
+      .reduce(
+        (acc, n) => {
+          const total =
+            n.macmon_info?.memory?.ram_total ?? n.system_info?.memory ?? 0;
+          const used = n.macmon_info?.memory?.ram_usage ?? 0;
+          return { used: acc.used + used, total: acc.total + total };
+        },
+        { used: 0, total: 0 },
+      );
  });
 </script>

+{#snippet clusterWarnings()}
+  {#if tbBridgeCycles.length > 0 || macosVersionMismatch || (tb5WithoutRdma && !tb5InfoDismissed)}
+    <div class="absolute top-4 left-4 flex flex-col gap-2 z-40">
+      {#if tbBridgeCycles.length > 0}
+        {@const cycle = tbBridgeCycles[0]}
+        {@const serviceName = getTbBridgeServiceName(cycle)}
+        {@const disableCmd = `sudo networksetup -setnetworkserviceenabled "${serviceName}" off`}
+        <div class="group relative" role="alert">
+          <div
+            class="flex items-center gap-2 px-3 py-2 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm cursor-help"
+          >
+            <svg
+              class="w-5 h-5 text-yellow-400 flex-shrink-0"
+              fill="none"
+              viewBox="0 0 24 24"
+              stroke="currentColor"
+              stroke-width="2"
+            >
+              <path
+                stroke-linecap="round"
+                stroke-linejoin="round"
+                d={warningIconPath}
+              />
+            </svg>
+            <span class="text-sm font-mono text-yellow-200">
+              THUNDERBOLT BRIDGE CYCLE DETECTED
+            </span>
+          </div>
+
+          <!-- Tooltip on hover -->
+          <div
+            class="absolute top-full left-0 mt-2 w-80 p-3 rounded border border-yellow-500/30 bg-exo-dark-gray/95 backdrop-blur-sm opacity-0 invisible group-hover:opacity-100 group-hover:visible transition-all duration-200 z-50 shadow-lg"
+          >
+            <p class="text-xs text-white/80 mb-2">
+              A network routing cycle was detected between nodes connected via
+              Thunderbolt Bridge. This can cause connectivity issues.
+            </p>
+            <p class="text-xs text-white/60 mb-2">
+              <span class="text-yellow-300">Affected nodes:</span>
+              {cycle.map(getNodeName).join(" → ")}
+            </p>
+            <p class="text-xs text-white/60 mb-1">
+              <span class="text-yellow-300">To fix:</span> Disable the Thunderbolt
+              Bridge on one of the affected nodes:
+            </p>
+            <button
+              type="button"
+              onclick={() => copyToClipboard(disableCmd)}
+              class="w-full flex items-center gap-2 text-[10px] font-mono bg-exo-black/60 px-2 py-1.5 rounded text-exo-yellow break-all text-left hover:bg-exo-black/80 transition-colors cursor-pointer group/copy"
+              title="Click to copy"
+            >
+              <span class="flex-1">{disableCmd}</span>
+              <svg
+                class="w-3.5 h-3.5 flex-shrink-0 text-white/40 group-hover/copy:text-exo-yellow transition-colors"
+                fill="none"
+                viewBox="0 0 24 24"
+                stroke="currentColor"
+                stroke-width="2"
+              >
+                {#if copiedCommand}
+                  <path
+                    stroke-linecap="round"
+                    stroke-linejoin="round"
+                    d="M5 13l4 4L19 7"
+                  />
+                {:else}
+                  <path
+                    stroke-linecap="round"
+                    stroke-linejoin="round"
+                    d="M8 16H6a2 2 0 01-2-2V6a2 2 0 012-2h8a2 2 0 012 2v2m-6 12h8a2 2 0 002-2v-8a2 2 0 00-2-2h-8a2 2 0 00-2 2v8a2 2 0 002 2z"
+                  />
+                {/if}
+              </svg>
+            </button>
+          </div>
+        </div>
+      {/if}
+
+      {#if macosVersionMismatch}
+        <div class="group relative" role="alert">
+          <div
+            class="flex items-center gap-2 px-3 py-2 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm cursor-help"
+          >
+            <svg
+              class="w-5 h-5 text-yellow-400 flex-shrink-0"
+              fill="none"
+              viewBox="0 0 24 24"
+              stroke="currentColor"
+              stroke-width="2"
+            >
+              <path
+                stroke-linecap="round"
+                stroke-linejoin="round"
+                d={warningIconPath}
+              />
+            </svg>
+            <span class="text-sm font-mono text-yellow-200">
+              INCOMPATIBLE macOS VERSIONS
+            </span>
+          </div>
+
+          <!-- Tooltip on hover -->
+          <div
+            class="absolute top-full left-0 mt-2 w-80 p-3 rounded border border-yellow-500/30 bg-exo-dark-gray/95 backdrop-blur-sm opacity-0 invisible group-hover:opacity-100 group-hover:visible transition-all duration-200 z-50 shadow-lg"
+          >
+            <p class="text-xs text-white/80 mb-2">
+              Nodes in this cluster are running different macOS versions. This
+              may cause inference compatibility issues.
+            </p>
+            <div class="text-xs text-white/60 mb-2">
+              <span class="text-yellow-300">Node versions:</span>
+              {#each macosVersionMismatch as node}
+                <div class="ml-2">
+                  {node.friendlyName} — macOS {node.version} ({node.buildVersion})
+                </div>
+              {/each}
+            </div>
+            <p class="text-xs text-white/60">
+              <span class="text-yellow-300">Suggested action:</span> Update all nodes
+              to the same macOS version for best compatibility.
+            </p>
+          </div>
+        </div>
+      {/if}
+
+      {#if tb5WithoutRdma && !tb5InfoDismissed}
+        <div
+          class="flex items-center gap-2 px-3 py-2 rounded border border-blue-400/50 bg-blue-400/10 backdrop-blur-sm"
+          role="status"
+        >
+          <svg
+            class="w-5 h-5 text-blue-400 flex-shrink-0"
+            fill="none"
+            viewBox="0 0 24 24"
+            stroke="currentColor"
+            stroke-width="2"
+          >
+            <path
+              stroke-linecap="round"
+              stroke-linejoin="round"
+              d={infoIconPath}
+            />
+          </svg>
+          <span class="text-sm font-mono text-blue-200"> RDMA AVAILABLE </span>
+          <button
+            type="button"
+            onclick={() => (tb5InfoDismissed = true)}
+            class="ml-1 text-blue-300/60 hover:text-blue-200 transition-colors cursor-pointer"
+            title="Dismiss"
+          >
+            <svg
+              class="w-4 h-4"
+              fill="none"
+              viewBox="0 0 24 24"
+              stroke="currentColor"
+              stroke-width="2"
+            >
+              <path
+                stroke-linecap="round"
+                stroke-linejoin="round"
+                d="M6 18L18 6M6 6l12 12"
+              />
+            </svg>
+          </button>
+        </div>
+      {/if}
+    </div>
+  {/if}
+{/snippet}
+
+{#snippet clusterWarningsCompact()}
+  {#if tbBridgeCycles.length > 0 || macosVersionMismatch || (tb5WithoutRdma && !tb5InfoDismissed)}
+    <div class="absolute top-2 left-2 flex flex-col gap-1">
+      {#if tbBridgeCycles.length > 0}
+        <div
+          class="flex items-center gap-1.5 px-2 py-1 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm"
+          title="Thunderbolt Bridge cycle detected"
+        >
+          <svg
+            class="w-3.5 h-3.5 text-yellow-400"
+            fill="none"
+            viewBox="0 0 24 24"
+            stroke="currentColor"
+            stroke-width="2"
+          >
+            <path
+              stroke-linecap="round"
+              stroke-linejoin="round"
+              d={warningIconPath}
+            />
+          </svg>
+          <span class="text-[10px] font-mono text-yellow-200">TB CYCLE</span>
+        </div>
+      {/if}
+      {#if macosVersionMismatch}
+        <div
+          class="flex items-center gap-1.5 px-2 py-1 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm"
+          title="Incompatible macOS versions detected"
+        >
+          <svg
+            class="w-3.5 h-3.5 text-yellow-400"
+            fill="none"
+            viewBox="0 0 24 24"
+            stroke="currentColor"
+            stroke-width="2"
+          >
+            <path
+              stroke-linecap="round"
+              stroke-linejoin="round"
+              d={warningIconPath}
+            />
+          </svg>
+          <span class="text-[10px] font-mono text-yellow-200"
+            >macOS MISMATCH</span
+          >
+        </div>
+      {/if}
+      {#if tb5WithoutRdma && !tb5InfoDismissed}
+        <div
+          class="flex items-center gap-1.5 px-2 py-1 rounded border border-blue-400/50 bg-blue-400/10 backdrop-blur-sm"
+          title="Thunderbolt 5 detected — RDMA can be enabled for better performance"
+        >
+          <svg
+            class="w-3.5 h-3.5 text-blue-400"
+            fill="none"
+            viewBox="0 0 24 24"
+            stroke="currentColor"
+            stroke-width="2"
+          >
+            <path
+              stroke-linecap="round"
+              stroke-linejoin="round"
+              d={infoIconPath}
+            />
+          </svg>
+          <span class="text-[10px] font-mono text-blue-200">RDMA AVAILABLE</span
+          >
+        </div>
+      {/if}
+    </div>
+  {/if}
+{/snippet}
+
 <!-- Global event listeners for slider dragging -->
 <svelte:window
  onmousemove={handleSliderMouseMove}
@@ -1724,17 +2084,40 @@
            onNodeClick={togglePreviewNodeFilter}
          />

-          <!-- Thunderbolt Bridge Cycle Warning -->
-          {#if tbBridgeCycles.length > 0}
-            {@const cycle = tbBridgeCycles[0]}
-            {@const serviceName = getTbBridgeServiceName(cycle)}
-            {@const disableCmd = `sudo networksetup -setnetworkserviceenabled "${serviceName}" off`}
-            <div class="absolute top-4 left-4 group" role="alert">
-              <div
-                class="flex items-center gap-2 px-3 py-2 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm cursor-help"
+          {@render clusterWarnings()}
+
+          <!-- TB5 RDMA Available Info -->
+          {#if tb5WithoutRdma && !tb5InfoDismissed}
+            <div
+              class="absolute left-4 flex items-center gap-2 px-3 py-2 rounded border border-blue-400/50 bg-blue-400/10 backdrop-blur-sm"
+              class:top-16={tbBridgeCycles.length > 0}
+              class:top-4={tbBridgeCycles.length === 0}
+              role="status"
+            >
+              <svg
+                class="w-5 h-5 text-blue-400 flex-shrink-0"
+                fill="none"
+                viewBox="0 0 24 24"
+                stroke="currentColor"
+                stroke-width="2"
+              >
+                <path
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                  d="M13 16h-1v-4h-1m1-4h.01M21 12a9 9 0 11-18 0 9 9 0 0118 0z"
+                />
+              </svg>
+              <span class="text-sm font-mono text-blue-200">
+                RDMA AVAILABLE
+              </span>
+              <button
+                type="button"
+                onclick={() => (tb5InfoDismissed = true)}
+                class="ml-1 text-blue-300/60 hover:text-blue-200 transition-colors cursor-pointer"
+                title="Dismiss"
              >
                <svg
-                  class="w-5 h-5 text-yellow-400 flex-shrink-0"
+                  class="w-4 h-4"
                  fill="none"
                  viewBox="0 0 24 24"
                  stroke="currentColor"
@@ -1743,60 +2126,10 @@
                  <path
                    stroke-linecap="round"
                    stroke-linejoin="round"
-                    d="M12 9v2m0 4h.01m-6.938 4h13.856c1.54 0 2.502-1.667 1.732-3L13.732 4c-.77-1.333-2.694-1.333-3.464 0L3.34 16c-.77 1.333.192 3 1.732 3z"
+                    d="M6 18L18 6M6 6l12 12"
                  />
                </svg>
-                <span class="text-sm font-mono text-yellow-200">
-                  THUNDERBOLT BRIDGE CYCLE DETECTED
-                </span>
-              </div>
-
-              <!-- Tooltip on hover -->
-              <div
-                class="absolute top-full left-0 mt-2 w-80 p-3 rounded border border-yellow-500/30 bg-exo-dark-gray/95 backdrop-blur-sm opacity-0 invisible group-hover:opacity-100 group-hover:visible transition-all duration-200 z-50 shadow-lg"
-              >
-                <p class="text-xs text-white/80 mb-2">
-                  A network routing cycle was detected between nodes connected
-                  via Thunderbolt Bridge. This can cause connectivity issues.
-                </p>
-                <p class="text-xs text-white/60 mb-2">
-                  <span class="text-yellow-300">Affected nodes:</span>
-                  {cycle.map(getNodeName).join(" → ")}
-                </p>
-                <p class="text-xs text-white/60 mb-1">
-                  <span class="text-yellow-300">To fix:</span> Disable the Thunderbolt
-                  Bridge on one of the affected nodes:
-                </p>
-                <button
-                  type="button"
-                  onclick={() => copyToClipboard(disableCmd)}
-                  class="w-full flex items-center gap-2 text-[10px] font-mono bg-exo-black/60 px-2 py-1.5 rounded text-exo-yellow break-all text-left hover:bg-exo-black/80 transition-colors cursor-pointer group/copy"
-                  title="Click to copy"
-                >
-                  <span class="flex-1">{disableCmd}</span>
-                  <svg
-                    class="w-3.5 h-3.5 flex-shrink-0 text-white/40 group-hover/copy:text-exo-yellow transition-colors"
-                    fill="none"
-                    viewBox="0 0 24 24"
-                    stroke="currentColor"
-                    stroke-width="2"
-                  >
-                    {#if copiedCommand}
-                      <path
-                        stroke-linecap="round"
-                        stroke-linejoin="round"
-                        d="M5 13l4 4L19 7"
-                      />
-                    {:else}
-                      <path
-                        stroke-linecap="round"
-                        stroke-linejoin="round"
-                        d="M8 16H6a2 2 0 01-2-2V6a2 2 0 012-2h8a2 2 0 012 2v2m-6 12h8a2 2 0 002-2v-8a2 2 0 00-2-2h-8a2 2 0 00-2 2v8a2 2 0 002 2z"
-                      />
-                    {/if}
-                  </svg>
-                </button>
-              </div>
+              </button>
            </div>
          {/if}

@@ -1843,17 +2176,21 @@
              onNodeClick={togglePreviewNodeFilter}
            />

-            <!-- Thunderbolt Bridge Cycle Warning -->
-            {#if tbBridgeCycles.length > 0}
-              {@const cycle = tbBridgeCycles[0]}
-              {@const serviceName = getTbBridgeServiceName(cycle)}
-              {@const disableCmd = `sudo networksetup -setnetworkserviceenabled "${serviceName}" off`}
-              <div class="absolute top-4 left-4 group" role="alert">
+            {@render clusterWarnings()}
+
+            <!-- TB5 RDMA Available Info -->
+            {#if tb5WithoutRdma && !tb5InfoDismissed}
+              <div
+                class="absolute left-4 group"
+                class:top-16={tbBridgeCycles.length > 0}
+                class:top-4={tbBridgeCycles.length === 0}
+                role="status"
+              >
                <div
-                  class="flex items-center gap-2 px-3 py-2 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm cursor-help"
+                  class="flex items-center gap-2 px-3 py-2 rounded border border-blue-400/50 bg-blue-400/10 backdrop-blur-sm"
                >
                  <svg
-                    class="w-5 h-5 text-yellow-400 flex-shrink-0"
+                    class="w-5 h-5 text-blue-400 flex-shrink-0"
                    fill="none"
                    viewBox="0 0 24 24"
                    stroke="currentColor"
@@ -1862,60 +2199,62 @@
                    <path
                      stroke-linecap="round"
                      stroke-linejoin="round"
-                      d="M12 9v2m0 4h.01m-6.938 4h13.856c1.54 0 2.502-1.667 1.732-3L13.732 4c-.77-1.333-2.694-1.333-3.464 0L3.34 16c-.77 1.333.192 3 1.732 3z"
+                      d="M13 16h-1v-4h-1m1-4h.01M21 12a9 9 0 11-18 0 9 9 0 0118 0z"
                    />
                  </svg>
-                  <span class="text-sm font-mono text-yellow-200">
-                    THUNDERBOLT BRIDGE CYCLE DETECTED
+                  <span class="text-sm font-mono text-blue-200">
+                    RDMA AVAILABLE
                  </span>
-                </div>
-
-                <!-- Tooltip on hover -->
-                <div
-                  class="absolute top-full left-0 mt-2 w-80 p-3 rounded border border-yellow-500/30 bg-exo-dark-gray/95 backdrop-blur-sm opacity-0 invisible group-hover:opacity-100 group-hover:visible transition-all duration-200 z-50 shadow-lg"
-                >
-                  <p class="text-xs text-white/80 mb-2">
-                    A network routing cycle was detected between nodes connected
-                    via Thunderbolt Bridge. This can cause connectivity issues.
-                  </p>
-                  <p class="text-xs text-white/60 mb-2">
-                    <span class="text-yellow-300">Affected nodes:</span>
-                    {cycle.map(getNodeName).join(" → ")}
-                  </p>
-                  <p class="text-xs text-white/60 mb-1">
-                    <span class="text-yellow-300">To fix:</span> Disable the Thunderbolt
-                    Bridge on one of the affected nodes:
-                  </p>
                  <button
                    type="button"
-                    onclick={() => copyToClipboard(disableCmd)}
-                    class="w-full flex items-center gap-2 text-[10px] font-mono bg-exo-black/60 px-2 py-1.5 rounded text-exo-yellow break-all text-left hover:bg-exo-black/80 transition-colors cursor-pointer group/copy"
-                    title="Click to copy"
+                    onclick={() => (tb5InfoDismissed = true)}
+                    class="ml-1 text-blue-300/60 hover:text-blue-200 transition-colors cursor-pointer"
+                    title="Dismiss"
                  >
-                    <span class="flex-1">{disableCmd}</span>
                    <svg
-                      class="w-3.5 h-3.5 flex-shrink-0 text-white/40 group-hover/copy:text-exo-yellow transition-colors"
+                      class="w-4 h-4"
                      fill="none"
                      viewBox="0 0 24 24"
                      stroke="currentColor"
                      stroke-width="2"
                    >
-                      {#if copiedCommand}
-                        <path
-                          stroke-linecap="round"
-                          stroke-linejoin="round"
-                          d="M5 13l4 4L19 7"
-                        />
-                      {:else}
-                        <path
-                          stroke-linecap="round"
-                          stroke-linejoin="round"
-                          d="M8 16H6a2 2 0 01-2-2V6a2 2 0 012-2h8a2 2 0 012 2v2m-6 12h8a2 2 0 002-2v-8a2 2 0 00-2-2h-8a2 2 0 00-2 2v8a2 2 0 002 2z"
-                        />
-                      {/if}
+                      <path
+                        stroke-linecap="round"
+                        stroke-linejoin="round"
+                        d="M6 18L18 6M6 6l12 12"
+                      />
                    </svg>
                  </button>
                </div>
+
+                <!-- Tooltip on hover -->
+                <div
+                  class="absolute top-full left-0 mt-2 w-80 p-3 rounded border border-blue-400/30 bg-exo-dark-gray/95 backdrop-blur-sm opacity-0 invisible group-hover:opacity-100 group-hover:visible transition-all duration-200 z-50 shadow-lg"
+                >
+                  <p class="text-xs text-white/80 mb-2">
+                    Thunderbolt 5 hardware detected on multiple nodes. Enable
+                    RDMA for significantly faster inter-node communication.
+                  </p>
+                  <p class="text-xs text-white/60 mb-1.5">
+                    <span class="text-blue-300">To enable:</span>
+                  </p>
+                  <ol
+                    class="text-xs text-white/60 list-decimal list-inside space-y-0.5 mb-1.5"
+                  >
+                    <li>Connect nodes with TB5 cables</li>
+                    <li>Boot to Recovery (hold power 10s → Options)</li>
+                    <li>
+                      Run
+                      <code class="text-blue-300 bg-blue-400/10 px-1 rounded"
+                        >rdma_ctl enable</code
+                      >
+                    </li>
+                    <li>Reboot</li>
+                  </ol>
+                  <p class="text-xs text-white/40">
+                    Requires macOS 26.2+, TB5 cables, and matching OS versions.
+                  </p>
+                </div>
              </div>
            {/if}

@@ -1954,6 +2293,7 @@
                showHelperText={false}
                showModelSelector={true}
                modelTasks={modelTasks()}
+                modelCapabilities={modelCapabilities()}
              />
            </div>
          </div>
@@ -2733,6 +3073,7 @@
                placeholder="Ask anything"
                showModelSelector={true}
                modelTasks={modelTasks()}
+                modelCapabilities={modelCapabilities()}
              />
            </div>
          </div>
@@ -2760,7 +3101,7 @@
                  TOPOLOGY
                </div>
                <span class="text-xs text-white/70 tabular-nums"
-                  >{nodeCount} {nodeCount === 1 ? "NODE" : "NODES"}</span
+                  >{totalNodeCount} {totalNodeCount === 1 ? "NODE" : "NODES"}</span
                >
              </div>

@@ -2773,30 +3114,7 @@
                  onNodeClick={togglePreviewNodeFilter}
                />

-                <!-- Thunderbolt Bridge Cycle Warning (compact) -->
-                {#if tbBridgeCycles.length > 0}
-                  <div
-                    class="absolute top-2 left-2 flex items-center gap-1.5 px-2 py-1 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm"
-                    title="Thunderbolt Bridge cycle detected - click to view details"
-                  >
-                    <svg
-                      class="w-3.5 h-3.5 text-yellow-400"
-                      fill="none"
-                      viewBox="0 0 24 24"
-                      stroke="currentColor"
-                      stroke-width="2"
-                    >
-                      <path
-                        stroke-linecap="round"
-                        stroke-linejoin="round"
-                        d="M12 9v2m0 4h.01m-6.938 4h13.856c1.54 0 2.502-1.667 1.732-3L13.732 4c-.77-1.333-2.694-1.333-3.464 0L3.34 16c-.77 1.333.192 3 1.732 3z"
-                      />
-                    </svg>
-                    <span class="text-[10px] font-mono text-yellow-200"
-                      >TB CYCLE</span
-                    >
-                  </div>
-                {/if}
+                {@render clusterWarningsCompact()}
              </div>
            </button>

@@ -3252,11 +3570,17 @@
  {models}
  {selectedModelId}
  favorites={favoritesSet}
+  {recentModelIds}
+  hasRecents={showRecentsTab}
  existingModelIds={new Set(models.map((m) => m.id))}
  canModelFit={(modelId) => {
    const model = models.find((m) => m.id === modelId);
    return model ? hasEnoughMemory(model) : false;
  }}
+  getModelFitStatus={(modelId): ModelMemoryFitStatus => {
+    const model = models.find((m) => m.id === modelId);
+    return model ? getModelMemoryFitStatus(model) : "too_large";
+  }}
  onSelect={handleModelPickerSelect}
  onClose={() => (isModelPickerOpen = false)}
  onToggleFavorite={toggleFavorite}
@@ -83,6 +83,9 @@
          _module.args.pkgs = import inputs.nixpkgs {
            inherit system;
            config.allowUnfreePredicate = pkg: (pkg.pname or "") == "metal-toolchain";
+            overlays = [
+              (import ./nix/apple-sdk-overlay.nix)
+            ];
          };
          treefmt = {
            projectRootFile = "flake.nix";
@@ -0,0 +1,18 @@
+# Overlay that builds apple-sdk with a custom versions.json (for SDK 26.2).
+# The upstream nixpkgs package reads versions.json at eval time via a relative
+# path, so we can't override it through callPackage args. Instead, we copy
+# the upstream source and patch the one file.
+final: _prev:
+let
+  upstreamSrc = final.path + "/pkgs/by-name/ap/apple-sdk";
+  patchedSrc = final.runCommandLocal "apple-sdk-src-patched" { } ''
+    cp -r ${upstreamSrc} $out
+    chmod -R u+w $out
+    cp ${./apple-sdk/metadata/versions.json} $out/metadata/versions.json
+  '';
+in
+{
+  apple-sdk_26 = final.callPackage (patchedSrc + "/package.nix") {
+    darwinSdkMajorVersion = "26";
+  };
+}
@@ -0,0 +1,26 @@
+{
+  "14": {
+    "urls": [
+      "https://swcdn.apple.com/content/downloads/14/48/052-59890-A_I0F5YGAY0Y/p9n40hio7892gou31o1v031ng6fnm9sb3c/CLTools_macOSNMOS_SDK.pkg",
+      "https://web.archive.org/web/20250211001355/https://swcdn.apple.com/content/downloads/14/48/052-59890-A_I0F5YGAY0Y/p9n40hio7892gou31o1v031ng6fnm9sb3c/CLTools_macOSNMOS_SDK.pkg"
+    ],
+    "version": "14.4",
+    "hash": "sha256-QozDiwY0Czc0g45vPD7G4v4Ra+3DujCJbSads3fJjjM="
+  },
+  "15": {
+    "urls": [
+      "https://swcdn.apple.com/content/downloads/52/01/082-41241-A_0747ZN8FHV/dectd075r63pppkkzsb75qk61s0lfee22j/CLTools_macOSNMOS_SDK.pkg",
+      "https://web.archive.org/web/20250530132510/https://swcdn.apple.com/content/downloads/52/01/082-41241-A_0747ZN8FHV/dectd075r63pppkkzsb75qk61s0lfee22j/CLTools_macOSNMOS_SDK.pkg"
+    ],
+    "version": "15.5",
+    "hash": "sha256-HBiSJuw1XBUK5R/8Sj65c3rftSEvQl/O9ZZVp/g1Amo="
+  },
+  "26": {
+    "urls": [
+      "https://swcdn.apple.com/content/downloads/60/22/089-71960-A_W8BL1RUJJ6/5zkyplomhk1cm7z6xja2ktgapnhhti6wwd/CLTools_macOSNMOS_SDK.pkg",
+      "https://web.archive.org/web/20250915230423/https://swcdn.apple.com/content/downloads/60/22/089-71960-A_W8BL1RUJJ6/5zkyplomhk1cm7z6xja2ktgapnhhti6wwd/CLTools_macOSNMOS_SDK.pkg"
+    ],
+    "version": "26.2",
+    "hash": "sha256-hXRlMieVv0smna5uiWRwq87IWOaPWtAjAldbi+wQXcw="
+  }
+}
@@ -41,7 +41,7 @@ let

  mlx = stdenv.mkDerivation rec {
    pname = "mlx";
-    version = let v = "0.30.5"; in
+    version = let v = "0.30.6"; in
      assert v == uvLockMlxVersion || throw "MLX version mismatch: nix/mlx.nix has ${v} but uv.lock has ${uvLockMlxVersion}. Update both the version and hash in nix/mlx.nix.";
      v;
    pyproject = true;
@@ -50,7 +50,7 @@ let
      owner = "ml-explore";
      repo = "mlx";
      tag = "v${version}";
-      hash = "sha256-SV/3MXt+SuJ69XfLfXycold6KgtXSM7OE0KwMSNw+eE=";
+      hash = "sha256-avD5EGhwgmPdXLAyQSqTO6AXk/W3ziH+f6AetjK3Sdo=";
    };

    patches = [
@@ -85,6 +85,7 @@ let
        (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_JSON" "${nlohmann_json.src}")
        (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_NANOBIND" "${nanobind}")
        (lib.cmakeBool "FETCHCONTENT_FULLY_DISCONNECTED" true)
+        (lib.cmakeBool "MLX_BUILD_CPU" true)
        (lib.cmakeBool "MLX_BUILD_METAL" true)
        (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_METAL_CPP" "${metal_cpp}")
        (lib.cmakeOptionType "string" "CMAKE_OSX_DEPLOYMENT_TARGET" "${apple-sdk_26.version}")
@@ -17,8 +17,8 @@ dependencies = [
    "loguru>=0.7.3",
    "exo_pyo3_bindings", # rust bindings
    "anyio==4.11.0",
-    "mlx==0.30.5; sys_platform == 'darwin'",
-    "mlx[cpu]==0.30.5; sys_platform == 'linux'",
+    "mlx==0.30.6; sys_platform == 'darwin'",
+    "mlx[cpu]==0.30.6; sys_platform == 'linux'",
    "mlx-lm==0.30.6",
    "tiktoken>=0.12.0", # required for kimi k2 tokenizer
    "hypercorn>=0.18.0",
@@ -28,6 +28,8 @@ dependencies = [
    "pillow>=11.0,<12.0", # compatibility with mflux
    "mflux==0.15.5",
    "python-multipart>=0.0.21",
+    "msgspec>=0.19.0",
+    "zstandard>=0.23.0",
 ]

 [project.scripts]
@@ -130,7 +132,7 @@ markers = [
 env = [
  "EXO_TESTS=1"
 ]
-addopts = "-m 'not slow'"
+addopts = "-m 'not slow' --ignore=tests/start_distributed_test.py"
 filterwarnings = [
    "ignore:builtin type Swig:DeprecationWarning",
 ]
@@ -14,7 +14,9 @@

      # Override overlay to inject Nix-built components
      exoOverlay = final: prev: {
-        # Replace workspace exo_pyo3_bindings with Nix-built wheel
+        # Replace workspace exo_pyo3_bindings with Nix-built wheel.
+        # Preserve passthru so mkVirtualEnv can resolve dependency groups.
+        # Copy .pyi stub + py.typed marker so basedpyright can find the types.
        exo-pyo3-bindings = pkgs.stdenv.mkDerivation {
          pname = "exo-pyo3-bindings";
          version = "0.1.0";
@@ -22,6 +24,12 @@
          # Install from pre-built wheel
          nativeBuildInputs = [ final.pyprojectWheelHook ];
          dontStrip = true;
+          passthru = prev.exo-pyo3-bindings.passthru or { };
+          postInstall = ''
+            local siteDir=$out/${final.python.sitePackages}/exo_pyo3_bindings
+            cp ${inputs.self}/rust/exo_pyo3_bindings/exo_pyo3_bindings.pyi $siteDir/
+            touch $siteDir/py.typed
+          '';
        };
      };

@@ -29,17 +37,32 @@

      # Overlay to provide build systems and custom packages
      buildSystemsOverlay = final: prev: {
-        # Use our pure Nix-built MLX with Metal support
-        mlx = self'.packages.mlx;
-
        # mlx-lm is a git dependency that needs setuptools
        mlx-lm = prev.mlx-lm.overrideAttrs (old: {
          nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [
            final.setuptools
          ];
        });
+      } // lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin {
+        # Use our pure Nix-built MLX with Metal support (macOS only)
+        mlx = self'.packages.mlx;
      };

+      # Additional overlay for Linux-specific fixes (type checking env).
+      # Native wheels have shared lib dependencies we don't need at type-check time.
+      linuxOverlay = final: prev:
+        let
+          ignoreMissing = drv: drv.overrideAttrs { autoPatchelfIgnoreMissingDeps = [ "*" ]; };
+          nvidiaPackages = lib.filterAttrs (name: _: lib.hasPrefix "nvidia-" name) prev;
+        in
+        lib.optionalAttrs pkgs.stdenv.hostPlatform.isLinux (
+          (lib.mapAttrs (_: ignoreMissing) nvidiaPackages) // {
+            mlx = ignoreMissing prev.mlx;
+            torch = ignoreMissing prev.torch;
+            triton = ignoreMissing prev.triton;
+          }
+        );
+
      pythonSet = (pkgs.callPackage inputs.pyproject-nix.build.packages {
        inherit python;
      }).overrideScope (
@@ -48,6 +71,7 @@
          overlay
          exoOverlay
          buildSystemsOverlay
+          linuxOverlay
        ]
      );
      exoVenv = pythonSet.mkVirtualEnv "exo-env" workspace.deps.default;
@@ -118,6 +142,21 @@
          ${pkgs.ruff}/bin/ruff check ${inputs.self}
          touch $out
        '';
+
+        # Hermetic basedpyright type checking
+        typecheck = pkgs.runCommand "typecheck"
+          {
+            nativeBuildInputs = [
+              testVenv
+              pkgs.basedpyright
+            ];
+          }
+          ''
+            cd ${inputs.self}
+            export HOME=$TMPDIR
+            basedpyright --pythonpath ${testVenv}/bin/python
+            touch $out
+          '';
      };
    };
 }
@@ -0,0 +1,12 @@
+model_id = "mlx-community/Step-3.5-Flash-4bit"
+n_layers = 45
+hidden_size = 4096
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "step"
+quantization = "4bit"
+base_model = "Step 3.5 Flash"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 114572190076
@@ -0,0 +1,12 @@
+model_id = "mlx-community/Step-3.5-Flash-6bit"
+n_layers = 45
+hidden_size = 4096
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "step"
+quantization = "6bit"
+base_model = "Step 3.5 Flash"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 159039627774
@@ -0,0 +1,12 @@
+model_id = "mlx-community/Step-3.5-Flash-8Bit"
+n_layers = 45
+hidden_size = 4096
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "step"
+quantization = "8bit"
+base_model = "Step 3.5 Flash"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 209082699847
@@ -22,7 +22,6 @@ use pyo3::{Bound, Py, PyErr, PyResult, PyTraverseError, PyVisit, Python, pymetho
 use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pyclass_enum, gen_stub_pymethods};
 use std::net::IpAddr;
 use tokio::sync::{Mutex, mpsc, oneshot};
-use util::ext::VecExt as _;

 mod exception {
    use pyo3::types::PyTuple;
@@ -532,7 +531,9 @@ impl PyNetworkingHandle {
            .recv_many_py(limit)
            .allow_threads_py() // allow-threads-aware async call
            .await?
-            .map(|(t, d)| (t, d.pybytes())))
+            .into_iter()
+            .map(|(t, d)| (t, d.pybytes()))
+            .collect())
    }

    // TODO: rn this blocks main thread if anything else is awaiting the channel (bc its a mutex)
@@ -106,7 +106,6 @@ mod behaviour {
    use crate::{alias, discovery};
    use libp2p::swarm::NetworkBehaviour;
    use libp2p::{gossipsub, identity};
-    use std::time::Duration;

    /// Behavior of the Swarm which composes all desired behaviors:
    /// Right now its just [`discovery::Behaviour`] and [`gossipsub::Behaviour`].
@@ -134,7 +133,6 @@ mod behaviour {
        gossipsub::Behaviour::new(
            MessageAuthenticity::Signed(keypair.clone()),
            ConfigBuilder::default()
-                .publish_queue_duration(Duration::from_secs(15))
                .max_transmit_size(1024 * 1024)
                .validation_mode(ValidationMode::Strict)
                .build()
@@ -13,13 +13,3 @@ path = "src/lib.rs"
 workspace = true

 [dependencies]
-# macro dependencies
-extend = { workspace = true }
-
-# utility dependencies
-thiserror = { workspace = true }
-once_cell = { workspace = true }
-internment = { workspace = true }
-derive_more = { workspace = true }
-bon = { workspace = true }
-recursion = { workspace = true }
@@ -1,53 +1 @@
-//! TODO: crate documentation
-//!
-//! this is here as a placeholder documentation
-//!
-//!
-
-// enable Rust-unstable features for convenience
-#![feature(trait_alias)]
-#![feature(stmt_expr_attributes)]
-#![feature(type_alias_impl_trait)]
-#![feature(specialization)]
-#![feature(unboxed_closures)]
-#![feature(const_trait_impl)]
-#![feature(fn_traits)]
-
-pub mod nonempty;
 pub mod wakerdeque;
-
-pub(crate) mod private {
-    // sealed traits support
-    pub trait Sealed {}
-    impl<T: ?Sized> Sealed for T {}
-}
-
-/// Namespace for all the type/trait aliases used by this crate.
-pub(crate) mod alias {}
-
-/// Namespace for crate-wide extension traits/methods
-pub mod ext {
-    use extend::ext;
-
-    #[ext(pub, name = BoxedSliceExt)]
-    impl<T> Box<[T]> {
-        #[inline]
-        fn map<B, F>(self, f: F) -> Box<[B]>
-        where
-            F: FnMut(T) -> B,
-        {
-            self.into_iter().map(f).collect()
-        }
-    }
-
-    #[ext(pub, name = VecExt)]
-    impl<T> Vec<T> {
-        #[inline]
-        fn map<B, F>(self, f: F) -> Vec<B>
-        where
-            F: FnMut(T) -> B,
-        {
-            self.into_iter().map(f).collect()
-        }
-    }
-}
@@ -1,138 +0,0 @@
-use std::slice::SliceIndex;
-use std::{ops, slice};
-use thiserror::Error;
-
-#[derive(Error, Debug)]
-#[error("Cannot create to `NonemptyArray` because the supplied slice is empty")]
-pub struct EmptySliceError;
-
-/// A pointer to a non-empty fixed-size slice allocated on the heap.
-#[derive(Clone, Debug, PartialEq, Eq, Hash, PartialOrd, Ord)]
-#[repr(transparent)]
-pub struct NonemptyArray<T>(Box<[T]>);
-
-#[allow(clippy::arbitrary_source_item_ordering)]
-impl<T> NonemptyArray<T> {
-    #[inline]
-    pub fn singleton(value: T) -> Self {
-        Self(Box::new([value]))
-    }
-
-    #[allow(clippy::missing_errors_doc)]
-    #[inline]
-    pub fn try_from_boxed_slice<S: Into<Box<[T]>>>(
-        boxed_slice: S,
-    ) -> Result<Self, EmptySliceError> {
-        let boxed_slice = boxed_slice.into();
-        if boxed_slice.is_empty() {
-            Err(EmptySliceError)
-        } else {
-            Ok(Self(boxed_slice))
-        }
-    }
-
-    #[must_use]
-    #[inline]
-    pub fn into_boxed_slice(self) -> Box<[T]> {
-        self.0
-    }
-
-    #[must_use]
-    #[inline]
-    pub fn to_vec(&self) -> Vec<T>
-    where
-        T: Clone,
-    {
-        self.0.to_vec()
-    }
-
-    #[must_use]
-    #[inline]
-    pub const fn as_slice(&self) -> &[T] {
-        &self.0
-    }
-
-    #[allow(clippy::indexing_slicing)]
-    #[must_use]
-    #[inline]
-    pub fn first(&self) -> &T {
-        &self.0[0]
-    }
-
-    #[allow(clippy::indexing_slicing, clippy::arithmetic_side_effects)]
-    #[must_use]
-    #[inline]
-    pub fn last(&self) -> &T {
-        &self.0[self.0.len() - 1]
-    }
-
-    #[must_use]
-    #[inline]
-    pub fn get<I>(&self, index: I) -> Option<&I::Output>
-    where
-        I: SliceIndex<[T]>,
-    {
-        self.0.get(index)
-    }
-
-    #[allow(clippy::len_without_is_empty)]
-    #[must_use]
-    #[inline]
-    pub const fn len(&self) -> usize {
-        self.0.len()
-    }
-
-    #[allow(clippy::iter_without_into_iter)]
-    #[inline]
-    pub fn iter(&self) -> slice::Iter<'_, T> {
-        self.0.iter()
-    }
-
-    #[allow(clippy::iter_without_into_iter)]
-    #[inline]
-    pub fn iter_mut(&mut self) -> slice::IterMut<'_, T> {
-        self.0.iter_mut()
-    }
-
-    #[inline]
-    #[must_use]
-    pub fn map<U, F: FnMut(T) -> U>(self, f: F) -> NonemptyArray<U> {
-        NonemptyArray(self.0.into_iter().map(f).collect())
-    }
-}
-
-impl<T> From<NonemptyArray<T>> for Box<[T]> {
-    #[inline]
-    fn from(value: NonemptyArray<T>) -> Self {
-        value.into_boxed_slice()
-    }
-}
-
-impl<T> ops::Index<usize> for NonemptyArray<T> {
-    type Output = T;
-
-    #[inline]
-    fn index(&self, index: usize) -> &Self::Output {
-        self.0.index(index)
-    }
-}
-
-impl<T> IntoIterator for NonemptyArray<T> {
-    type Item = T;
-    type IntoIter = std::vec::IntoIter<T>;
-
-    #[inline]
-    fn into_iter(self) -> Self::IntoIter {
-        self.into_boxed_slice().into_vec().into_iter()
-    }
-}
-
-impl<'a, T> IntoIterator for &'a NonemptyArray<T> {
-    type Item = &'a T;
-    type IntoIter = slice::Iter<'a, T>;
-
-    #[inline]
-    fn into_iter(self) -> Self::IntoIter {
-        self.iter()
-    }
-}
@@ -14,6 +14,7 @@ from exo.download.download_utils import (
    map_repo_download_progress_to_download_progress_data,
 )
 from exo.download.shard_downloader import ShardDownloader
+from exo.shared.constants import EXO_MODELS_DIR
 from exo.shared.models.model_cards import ModelId
 from exo.shared.types.commands import (
    CancelDownload,
@@ -56,8 +57,54 @@ class DownloadCoordinator:
    event_receiver: Receiver[Event] = field(init=False)
    _tg: TaskGroup = field(init=False, default_factory=anyio.create_task_group)

+    # Per-model throttle for download progress events
+    _last_progress_time: dict[ModelId, float] = field(default_factory=dict)
+
    def __post_init__(self) -> None:
        self.event_sender, self.event_receiver = channel[Event]()
+        self.shard_downloader.on_progress(self._download_progress_callback)
+
+    def _model_dir(self, model_id: ModelId) -> str:
+        return str(EXO_MODELS_DIR / model_id.normalize())
+
+    async def _download_progress_callback(
+        self, callback_shard: ShardMetadata, progress: RepoDownloadProgress
+    ) -> None:
+        model_id = callback_shard.model_card.model_id
+        throttle_interval_secs = 1.0
+
+        if progress.status == "complete":
+            completed = DownloadCompleted(
+                shard_metadata=callback_shard,
+                node_id=self.node_id,
+                total_bytes=progress.total_bytes,
+                model_directory=self._model_dir(model_id),
+            )
+            self.download_status[model_id] = completed
+            await self.event_sender.send(
+                NodeDownloadProgress(download_progress=completed)
+            )
+            if model_id in self.active_downloads:
+                del self.active_downloads[model_id]
+            self._last_progress_time.pop(model_id, None)
+        elif (
+            progress.status == "in_progress"
+            and current_time() - self._last_progress_time.get(model_id, 0.0)
+            > throttle_interval_secs
+        ):
+            ongoing = DownloadOngoing(
+                node_id=self.node_id,
+                shard_metadata=callback_shard,
+                download_progress=map_repo_download_progress_to_download_progress_data(
+                    progress
+                ),
+                model_directory=self._model_dir(model_id),
+            )
+            self.download_status[model_id] = ongoing
+            await self.event_sender.send(
+                NodeDownloadProgress(download_progress=ongoing)
+            )
+            self._last_progress_time[model_id] = current_time()

    async def run(self) -> None:
        logger.info("Starting DownloadCoordinator")
@@ -119,17 +166,21 @@ class DownloadCoordinator:
    async def _start_download(self, shard: ShardMetadata) -> None:
        model_id = shard.model_card.model_id

-        # Check if already downloading or complete
+        # Check if already downloading, complete, or recently failed
        if model_id in self.download_status:
            status = self.download_status[model_id]
-            if isinstance(status, (DownloadOngoing, DownloadCompleted)):
+            if isinstance(status, (DownloadOngoing, DownloadCompleted, DownloadFailed)):
                logger.debug(
-                    f"Download for {model_id} already in progress or complete, skipping"
+                    f"Download for {model_id} already in progress, complete, or failed, skipping"
                )
                return

        # Emit pending status
-        progress = DownloadPending(shard_metadata=shard, node_id=self.node_id)
+        progress = DownloadPending(
+            shard_metadata=shard,
+            node_id=self.node_id,
+            model_directory=self._model_dir(model_id),
+        )
        self.download_status[model_id] = progress
        await self.event_sender.send(NodeDownloadProgress(download_progress=progress))

@@ -143,6 +194,7 @@ class DownloadCoordinator:
                shard_metadata=shard,
                node_id=self.node_id,
                total_bytes=initial_progress.total_bytes,
+                model_directory=self._model_dir(model_id),
            )
            self.download_status[model_id] = completed
            await self.event_sender.send(
@@ -165,50 +217,11 @@ class DownloadCoordinator:
            download_progress=map_repo_download_progress_to_download_progress_data(
                initial_progress
            ),
+            model_directory=self._model_dir(model_id),
        )
        self.download_status[model_id] = status
        self.event_sender.send_nowait(NodeDownloadProgress(download_progress=status))

-        last_progress_time = 0.0
-        throttle_interval_secs = 1.0
-
-        async def download_progress_callback(
-            callback_shard: ShardMetadata, progress: RepoDownloadProgress
-        ) -> None:
-            nonlocal last_progress_time
-
-            if progress.status == "complete":
-                completed = DownloadCompleted(
-                    shard_metadata=callback_shard,
-                    node_id=self.node_id,
-                    total_bytes=progress.total_bytes,
-                )
-                self.download_status[callback_shard.model_card.model_id] = completed
-                await self.event_sender.send(
-                    NodeDownloadProgress(download_progress=completed)
-                )
-                # Clean up active download tracking
-                if callback_shard.model_card.model_id in self.active_downloads:
-                    del self.active_downloads[callback_shard.model_card.model_id]
-            elif (
-                progress.status == "in_progress"
-                and current_time() - last_progress_time > throttle_interval_secs
-            ):
-                ongoing = DownloadOngoing(
-                    node_id=self.node_id,
-                    shard_metadata=callback_shard,
-                    download_progress=map_repo_download_progress_to_download_progress_data(
-                        progress
-                    ),
-                )
-                self.download_status[callback_shard.model_card.model_id] = ongoing
-                await self.event_sender.send(
-                    NodeDownloadProgress(download_progress=ongoing)
-                )
-                last_progress_time = current_time()
-
-        self.shard_downloader.on_progress(download_progress_callback)
-
        async def download_wrapper() -> None:
            try:
                await self.shard_downloader.ensure_shard(shard)
@@ -218,6 +231,7 @@ class DownloadCoordinator:
                    shard_metadata=shard,
                    node_id=self.node_id,
                    error_message=str(e),
+                    model_directory=self._model_dir(model_id),
                )
                self.download_status[model_id] = failed
                await self.event_sender.send(
@@ -252,6 +266,7 @@ class DownloadCoordinator:
            pending = DownloadPending(
                shard_metadata=current_status.shard_metadata,
                node_id=self.node_id,
+                model_directory=self._model_dir(model_id),
            )
            await self.event_sender.send(
                NodeDownloadProgress(download_progress=pending)
@@ -283,16 +298,29 @@ class DownloadCoordinator:
                    _,
                    progress,
                ) in self.shard_downloader.get_shard_download_status():
+                    model_id = progress.shard.model_card.model_id
+
+                    # Active downloads emit progress via the callback — don't overwrite
+                    if model_id in self.active_downloads:
+                        continue
+
                    if progress.status == "complete":
                        status: DownloadProgress = DownloadCompleted(
                            node_id=self.node_id,
                            shard_metadata=progress.shard,
                            total_bytes=progress.total_bytes,
+                            model_directory=self._model_dir(
+                                progress.shard.model_card.model_id
+                            ),
                        )
                    elif progress.status in ["in_progress", "not_started"]:
                        if progress.downloaded_bytes_this_session.in_bytes == 0:
                            status = DownloadPending(
-                                node_id=self.node_id, shard_metadata=progress.shard
+                                node_id=self.node_id,
+                                shard_metadata=progress.shard,
+                                model_directory=self._model_dir(
+                                    progress.shard.model_card.model_id
+                                ),
                            )
                        else:
                            status = DownloadOngoing(
@@ -301,6 +329,9 @@ class DownloadCoordinator:
                                download_progress=map_repo_download_progress_to_download_progress_data(
                                    progress
                                ),
+                                model_directory=self._model_dir(
+                                    progress.shard.model_card.model_id
+                                ),
                            )
                    else:
                        continue
@@ -251,7 +251,8 @@ class Node:
 def main():
    args = Args.parse()
    soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
-    resource.setrlimit(resource.RLIMIT_NOFILE, (max(soft, 65535), hard))
+    target = min(max(soft, 65535), hard)
+    resource.setrlimit(resource.RLIMIT_NOFILE, (target, hard))

    mp.set_start_method("spawn")
    # TODO: Refactor the current verbosity system
@@ -17,6 +17,7 @@ from exo.shared.types.api import (
    LogprobsContentItem,
    StreamingChoiceResponse,
    ToolCall,
+    Usage,
 )
 from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
 from exo.shared.types.common import CommandId
@@ -79,6 +80,7 @@ def chat_request_to_text_generation(
        seed=request.seed,
        stream=request.stream,
        tools=request.tools,
+        enable_thinking=request.enable_thinking,
        chat_template_messages=chat_template_messages
        if chat_template_messages
        else None,
@@ -124,6 +126,8 @@ async def generate_chat_stream(
    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
 ) -> AsyncGenerator[str, None]:
    """Generate Chat Completions API streaming events from chunks."""
+    last_usage: Usage | None = None
+
    async for chunk in chunk_stream:
        if isinstance(chunk, ErrorChunk):
            error_response = ErrorResponse(
@@ -137,6 +141,8 @@ async def generate_chat_stream(
            yield "data: [DONE]\n\n"
            return

+        last_usage = chunk.usage or last_usage
+
        if isinstance(chunk, ToolCallChunk):
            tool_call_deltas = [
                ToolCall(
@@ -160,12 +166,15 @@ async def generate_chat_stream(
                        finish_reason="tool_calls",
                    )
                ],
+                usage=last_usage,
            )
            yield f"data: {tool_response.model_dump_json()}\n\n"
            yield "data: [DONE]\n\n"
            return

        chunk_response = chunk_to_response(chunk, command_id)
+        if chunk.finish_reason is not None:
+            chunk_response = chunk_response.model_copy(update={"usage": last_usage})
        yield f"data: {chunk_response.model_dump_json()}\n\n"

        if chunk.finish_reason is not None:
@@ -175,7 +184,9 @@ async def generate_chat_stream(
 async def collect_chat_response(
    command_id: CommandId,
    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
-) -> ChatCompletionResponse:
+) -> AsyncGenerator[str]:
+    # This is an AsyncGenerator[str] rather than returning a ChatCompletionReponse because
+    # FastAPI handles the cancellation better but wouldn't auto-serialize for some reason
    """Collect all token chunks and return a single ChatCompletionResponse."""
    text_parts: list[str] = []
    tool_calls: list[ToolCall] = []
@@ -183,6 +194,7 @@ async def collect_chat_response(
    model: str | None = None
    finish_reason: FinishReason | None = None
    error_message: str | None = None
+    last_usage: Usage | None = None

    async for chunk in chunk_stream:
        if isinstance(chunk, ErrorChunk):
@@ -192,6 +204,8 @@ async def collect_chat_response(
        if model is None:
            model = chunk.model

+        last_usage = chunk.usage or last_usage
+
        if isinstance(chunk, TokenChunk):
            text_parts.append(chunk.text)
            if chunk.logprob is not None:
@@ -222,7 +236,7 @@ async def collect_chat_response(
    combined_text = "".join(text_parts)
    assert model is not None

-    return ChatCompletionResponse(
+    yield ChatCompletionResponse(
        id=command_id,
        created=int(time.time()),
        model=model,
@@ -240,4 +254,6 @@ async def collect_chat_response(
                finish_reason=finish_reason,
            )
        ],
-    )
+        usage=last_usage,
+    ).model_dump_json()
+    return
@@ -4,7 +4,7 @@ import json
 from collections.abc import AsyncGenerator
 from typing import Any

-from exo.shared.types.api import FinishReason
+from exo.shared.types.api import FinishReason, Usage
 from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
 from exo.shared.types.claude_api import (
    ClaudeContentBlock,
@@ -161,12 +161,14 @@ async def collect_claude_response(
    command_id: CommandId,
    model: str,
    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
-) -> ClaudeMessagesResponse:
+) -> AsyncGenerator[str]:
+    # This is an AsyncGenerator[str] rather than returning a ChatCompletionReponse because
+    # FastAPI handles the cancellation better but wouldn't auto-serialize for some reason
    """Collect all token chunks and return a single ClaudeMessagesResponse."""
    text_parts: list[str] = []
    tool_use_blocks: list[ClaudeToolUseBlock] = []
    stop_reason: ClaudeStopReason | None = None
-    last_stats = None
+    last_usage: Usage | None = None
    error_message: str | None = None

    async for chunk in chunk_stream:
@@ -174,6 +176,8 @@ async def collect_claude_response(
            error_message = chunk.error_message or "Internal server error"
            break

+        last_usage = chunk.usage or last_usage
+
        if isinstance(chunk, ToolCallChunk):
            for tool in chunk.tool_calls:
                tool_use_blocks.append(
@@ -183,12 +187,10 @@ async def collect_claude_response(
                        input=json.loads(tool.arguments),  # pyright: ignore[reportAny]
                    )
                )
-            last_stats = chunk.stats or last_stats
            stop_reason = "tool_use"
            continue

        text_parts.append(chunk.text)
-        last_stats = chunk.stats or last_stats

        if chunk.finish_reason is not None:
            stop_reason = finish_reason_to_claude_stop_reason(chunk.finish_reason)
@@ -208,11 +210,11 @@ async def collect_claude_response(
    if not content:
        content.append(ClaudeTextBlock(text=""))

-    # Use actual usage data from stats if available
-    input_tokens = last_stats.prompt_tokens if last_stats else 0
-    output_tokens = last_stats.generation_tokens if last_stats else 0
+    # Use actual usage data if available
+    input_tokens = last_usage.prompt_tokens if last_usage else 0
+    output_tokens = last_usage.completion_tokens if last_usage else 0

-    return ClaudeMessagesResponse(
+    yield ClaudeMessagesResponse(
        id=f"msg_{command_id}",
        model=model,
        content=content,
@@ -221,7 +223,8 @@ async def collect_claude_response(
            input_tokens=input_tokens,
            output_tokens=output_tokens,
        ),
-    )
+    ).model_dump_json()
+    return


 async def generate_claude_stream(
@@ -249,7 +252,7 @@ async def generate_claude_stream(

    output_tokens = 0
    stop_reason: ClaudeStopReason | None = None
-    last_stats = None
+    last_usage: Usage | None = None
    next_block_index = 1  # text block is 0, tool blocks start at 1

    async for chunk in chunk_stream:
@@ -257,8 +260,9 @@ async def generate_claude_stream(
            # Close text block and bail
            break

+        last_usage = chunk.usage or last_usage
+
        if isinstance(chunk, ToolCallChunk):
-            last_stats = chunk.stats or last_stats
            stop_reason = "tool_use"

            # Emit tool_use content blocks
@@ -290,7 +294,6 @@ async def generate_claude_stream(
            continue

        output_tokens += 1  # Count each chunk as one token
-        last_stats = chunk.stats or last_stats

        # content_block_delta
        delta_event = ClaudeContentBlockDeltaEvent(
@@ -302,9 +305,9 @@ async def generate_claude_stream(
        if chunk.finish_reason is not None:
            stop_reason = finish_reason_to_claude_stop_reason(chunk.finish_reason)

-    # Use actual token count from stats if available
-    if last_stats is not None:
-        output_tokens = last_stats.generation_tokens
+    # Use actual token count from usage if available
+    if last_usage is not None:
+        output_tokens = last_usage.completion_tokens

    # content_block_stop for text block
    block_stop = ClaudeContentBlockStopEvent(index=0)
@@ -4,6 +4,7 @@ from collections.abc import AsyncGenerator
 from itertools import count
 from typing import Any

+from exo.shared.types.api import Usage
 from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
 from exo.shared.types.common import CommandId
 from exo.shared.types.openai_responses import (
@@ -121,13 +122,15 @@ async def collect_responses_response(
    command_id: CommandId,
    model: str,
    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
-) -> ResponsesResponse:
+) -> AsyncGenerator[str]:
+    # This is an AsyncGenerator[str] rather than returning a ChatCompletionReponse because
+    # FastAPI handles the cancellation better but wouldn't auto-serialize for some reason
    """Collect all token chunks and return a single ResponsesResponse."""
    response_id = f"resp_{command_id}"
    item_id = f"item_{command_id}"
    accumulated_text = ""
    function_call_items: list[ResponseFunctionCallItem] = []
-    last_stats = None
+    last_usage: Usage | None = None
    error_message: str | None = None

    async for chunk in chunk_stream:
@@ -135,32 +138,32 @@ async def collect_responses_response(
            error_message = chunk.error_message or "Internal server error"
            break

+        last_usage = chunk.usage or last_usage
+
        if isinstance(chunk, ToolCallChunk):
            for tool in chunk.tool_calls:
                function_call_items.append(
                    ResponseFunctionCallItem(
-                        id=f"fc_{tool.id}",
-                        call_id=f"call_{tool.id}",
+                        id=tool.id,
+                        call_id=tool.id,
                        name=tool.name,
                        arguments=tool.arguments,
                    )
                )
-            last_stats = chunk.stats or last_stats
            continue

        accumulated_text += chunk.text
-        last_stats = chunk.stats or last_stats

    if error_message is not None:
        raise ValueError(error_message)

-    # Create usage from stats if available
+    # Create usage from usage data if available
    usage = None
-    if last_stats is not None:
+    if last_usage is not None:
        usage = ResponseUsage(
-            input_tokens=last_stats.prompt_tokens,
-            output_tokens=last_stats.generation_tokens,
-            total_tokens=last_stats.prompt_tokens + last_stats.generation_tokens,
+            input_tokens=last_usage.prompt_tokens,
+            output_tokens=last_usage.completion_tokens,
+            total_tokens=last_usage.total_tokens,
        )

    output: list[ResponseItem] = [
@@ -172,14 +175,15 @@ async def collect_responses_response(
    ]
    output.extend(function_call_items)

-    return ResponsesResponse(
+    yield ResponsesResponse(
        id=response_id,
        model=model,
        status="completed",
        output=output,
        output_text=accumulated_text,
        usage=usage,
-    )
+    ).model_dump_json()
+    return


 async def generate_responses_stream(
@@ -235,15 +239,16 @@ async def generate_responses_stream(

    accumulated_text = ""
    function_call_items: list[ResponseFunctionCallItem] = []
-    last_stats = None
+    last_usage: Usage | None = None
    next_output_index = 1  # message item is at 0

    async for chunk in chunk_stream:
        if isinstance(chunk, ErrorChunk):
            break

+        last_usage = chunk.usage or last_usage
+
        if isinstance(chunk, ToolCallChunk):
-            last_stats = chunk.stats or last_stats
            for tool in chunk.tool_calls:
                fc_id = f"fc_{tool.id}"
                call_id = f"call_{tool.id}"
@@ -302,7 +307,6 @@ async def generate_responses_stream(
            continue

        accumulated_text += chunk.text
-        last_stats = chunk.stats or last_stats

        # response.output_text.delta
        delta_event = ResponseTextDeltaEvent(
@@ -346,13 +350,13 @@ async def generate_responses_stream(
    )
    yield f"event: response.output_item.done\ndata: {item_done.model_dump_json()}\n\n"

-    # Create usage from stats if available
+    # Create usage from usage data if available
    usage = None
-    if last_stats is not None:
+    if last_usage is not None:
        usage = ResponseUsage(
-            input_tokens=last_stats.prompt_tokens,
-            output_tokens=last_stats.generation_tokens,
-            total_tokens=last_stats.prompt_tokens + last_stats.generation_tokens,
+            input_tokens=last_usage.prompt_tokens,
+            output_tokens=last_usage.completion_tokens,
+            total_tokens=last_usage.total_tokens,
        )

    # response.completed
@@ -3,7 +3,7 @@ import contextlib
 import json
 import random
 import time
-from collections.abc import AsyncGenerator, Awaitable, Callable
+from collections.abc import AsyncGenerator, Awaitable, Callable, Iterator
 from datetime import datetime, timezone
 from http import HTTPStatus
 from pathlib import Path
@@ -37,11 +37,13 @@ from exo.master.adapters.responses import (
    generate_responses_stream,
    responses_request_to_text_generation,
 )
+from exo.master.event_log import DiskEventLog
 from exo.master.image_store import ImageStore
 from exo.master.placement import place_instance as get_instance_placements
 from exo.shared.apply import apply
 from exo.shared.constants import (
    DASHBOARD_DIR,
+    EXO_EVENT_LOG_DIR,
    EXO_IMAGE_CACHE_DIR,
    EXO_MAX_CHUNK_SIZE,
    EXO_TRACING_CACHE_DIR,
@@ -83,6 +85,7 @@ from exo.shared.types.api import (
    ImageGenerationTaskParams,
    ImageListItem,
    ImageListResponse,
+    LiteNodeHeartbeatRequest,
    ModelList,
    ModelListModel,
    PlaceInstanceParams,
@@ -120,9 +123,11 @@ from exo.shared.types.commands import (
    ForwarderDownloadCommand,
    ImageEdits,
    ImageGeneration,
+    LiteNodeHeartbeat,
    PlaceInstance,
    SendInputChunk,
    StartDownload,
+    TaskCancelled,
    TaskFinished,
    TextGeneration,
 )
@@ -145,6 +150,9 @@ from exo.shared.types.worker.shards import Sharding
 from exo.utils.banner import print_startup_banner
 from exo.utils.channels import Receiver, Sender, channel
 from exo.utils.event_buffer import OrderedBuffer
+from exo.utils.info_gatherer.info_gatherer import LiteNodeRegistration
+
+_API_EVENT_LOG_DIR = EXO_EVENT_LOG_DIR / "api"


 def _format_to_content_type(image_format: Literal["png", "jpeg", "webp"] | None) -> str:
@@ -175,7 +183,7 @@ class API:
        election_receiver: Receiver[ElectionMessage],
    ) -> None:
        self.state = State()
-        self._event_log: list[Event] = []
+        self._event_log = DiskEventLog(_API_EVENT_LOG_DIR)
        self.command_sender = command_sender
        self.download_command_sender = download_command_sender
        self.global_event_receiver = global_event_receiver
@@ -223,6 +231,8 @@ class API:

    def reset(self, new_session_id: SessionId, result_clock: int):
        logger.info("Resetting API State")
+        self._event_log.close()
+        self._event_log = DiskEventLog(_API_EVENT_LOG_DIR)
        self.state = State()
        self.session_id = new_session_id
        self.event_buffer = OrderedBuffer[Event]()
@@ -289,13 +299,14 @@ class API:
        self.app.post("/v1/messages", response_model=None)(self.claude_messages)
        self.app.post("/v1/responses", response_model=None)(self.openai_responses)
        self.app.get("/state")(lambda: self.state)
-        self.app.get("/events")(lambda: self._event_log)
+        self.app.get("/events")(self.stream_events)
        self.app.post("/download/start")(self.start_download)
        self.app.delete("/download/{node_id}/{model_id:path}")(self.delete_download)
        self.app.get("/v1/traces")(self.list_traces)
        self.app.get("/v1/traces/{task_id}")(self.get_trace)
        self.app.get("/v1/traces/{task_id}/stats")(self.get_trace_stats)
        self.app.get("/v1/traces/{task_id}/raw")(self.get_trace_raw)
+        self.app.post("/v1/lite_node/heartbeat")(self.lite_node_heartbeat)

    async def place_instance(self, payload: PlaceInstanceParams):
        command = PlaceInstance(
@@ -358,6 +369,7 @@ class API:
                node_network=self.state.node_network,
                topology=self.state.topology,
                current_instances=self.state.instances,
+                node_identities=self.state.node_identities,
            )
        except ValueError as exc:
            raise HTTPException(status_code=400, detail=str(exc)) from exc
@@ -383,7 +395,14 @@ class API:
        previews: list[PlacementPreview] = []
        required_nodes = set(node_ids) if node_ids else None

-        if len(list(self.state.topology.list_nodes())) == 0:
+        full_node_count = sum(
+            1
+            for nid in self.state.topology.list_nodes()
+            if self.state.node_identities.get(nid) is None
+            or self.state.node_identities[nid].node_type != "lite"
+        )
+
+        if full_node_count == 0:
            return PlacementPreviewResponse(previews=[])

        try:
@@ -398,9 +417,7 @@ class API:
                instance_combinations.extend(
                    [
                        (sharding, instance_meta, i)
-                        for i in range(
-                            1, len(list(self.state.topology.list_nodes())) + 1
-                        )
+                        for i in range(1, full_node_count + 1)
                    ]
                )
        # TODO: PDD
@@ -420,6 +437,7 @@ class API:
                    topology=self.state.topology,
                    current_instances=self.state.instances,
                    required_nodes=required_nodes,
+                    node_identities=self.state.node_identities,
                )
            except ValueError as exc:
                if (model_card.model_id, sharding, instance_meta, 0) not in seen:
@@ -534,16 +552,14 @@ class API:
                        break

        except anyio.get_cancelled_exc_class():
-            # TODO: TaskCancelled
-            """
-            self.command_sender.send_nowait(
-                ForwarderCommand(origin=self.node_id, command=command)
-            )
-            """
+            command = TaskCancelled(cancelled_command_id=command_id)
+            with anyio.CancelScope(shield=True):
+                await self.command_sender.send(
+                    ForwarderCommand(origin=self.node_id, command=command)
+                )
            raise
        finally:
-            command = TaskFinished(finished_command_id=command_id)
-            await self._send(command)
+            await self._send(TaskFinished(finished_command_id=command_id))
            if command_id in self._text_generation_queues:
                del self._text_generation_queues[command_id]

@@ -638,11 +654,14 @@ class API:
                    "X-Accel-Buffering": "no",
                },
            )
-
-        return await collect_chat_response(
-            command.command_id,
-            self._token_chunk_stream(command.command_id),
-        )
+        else:
+            return StreamingResponse(
+                collect_chat_response(
+                    command.command_id,
+                    self._token_chunk_stream(command.command_id),
+                ),
+                media_type="application/json",
+            )

    async def bench_chat_completions(
        self, payload: BenchChatCompletionRequest
@@ -658,8 +677,7 @@ class API:
        command = TextGeneration(task_params=task_params)
        await self._send(command)

-        response = await self._collect_text_generation_with_stats(command.command_id)
-        return response
+        return await self._collect_text_generation_with_stats(command.command_id)

    async def _resolve_and_validate_text_model(self, model_id: ModelId) -> ModelId:
        """Validate a text model exists and return the resolved model ID.
@@ -694,6 +712,22 @@ class API:
            )
        return resolved_model

+    def stream_events(self) -> StreamingResponse:
+        def _generate_json_array(events: Iterator[Event]) -> Iterator[str]:
+            yield "["
+            first = True
+            for event in events:
+                if not first:
+                    yield ","
+                first = False
+                yield event.model_dump_json()
+            yield "]"
+
+        return StreamingResponse(
+            _generate_json_array(self._event_log.read_all()),
+            media_type="application/json",
+        )
+
    async def get_image(self, image_id: str) -> FileResponse:
        stored = self._image_store.get(Id(image_id))
        if stored is None:
@@ -861,6 +895,11 @@ class API:
                        del image_metadata[key]

        except anyio.get_cancelled_exc_class():
+            command = TaskCancelled(cancelled_command_id=command_id)
+            with anyio.CancelScope(shield=True):
+                await self.command_sender.send(
+                    ForwarderCommand(origin=self.node_id, command=command)
+                )
            raise
        finally:
            await self._send(TaskFinished(finished_command_id=command_id))
@@ -942,6 +981,11 @@ class API:

            return (images, stats if capture_stats else None)
        except anyio.get_cancelled_exc_class():
+            command = TaskCancelled(cancelled_command_id=command_id)
+            with anyio.CancelScope(shield=True):
+                await self.command_sender.send(
+                    ForwarderCommand(origin=self.node_id, command=command)
+                )
            raise
        finally:
            await self._send(TaskFinished(finished_command_id=command_id))
@@ -1199,12 +1243,15 @@ class API:
                    "X-Accel-Buffering": "no",
                },
            )
-
-        return await collect_claude_response(
-            command.command_id,
-            payload.model,
-            self._token_chunk_stream(command.command_id),
-        )
+        else:
+            return StreamingResponse(
+                collect_claude_response(
+                    command.command_id,
+                    payload.model,
+                    self._token_chunk_stream(command.command_id),
+                ),
+                media_type="application/json",
+            )

    async def openai_responses(
        self, payload: ResponsesRequest
@@ -1232,17 +1279,40 @@ class API:
                },
            )

-        return await collect_responses_response(
-            command.command_id,
-            payload.model,
-            self._token_chunk_stream(command.command_id),
+        else:
+            return StreamingResponse(
+                collect_responses_response(
+                    command.command_id,
+                    payload.model,
+                    self._token_chunk_stream(command.command_id),
+                ),
+                media_type="application/json",
+            )
+
+    async def lite_node_heartbeat(self, payload: LiteNodeHeartbeatRequest) -> JSONResponse:
+        info = LiteNodeRegistration(
+            model=payload.model,
+            chip=payload.chip,
+            os_version=payload.os_version,
+            friendly_name=payload.friendly_name,
+            ram_total=payload.ram_total,
+            ram_available=payload.ram_available,
        )
+        command = LiteNodeHeartbeat(
+            target_node_id=NodeId(payload.node_id),
+            info=info,
+        )
+        await self._send(command)
+        return JSONResponse({"status": "ok"})

    def _calculate_total_available_memory(self) -> Memory:
-        """Calculate total available memory across all nodes in bytes."""
+        """Calculate total available memory across all non-lite nodes in bytes."""
        total_available = Memory()

-        for memory in self.state.node_memory.values():
+        for node_id, memory in self.state.node_memory.items():
+            identity = self.state.node_identities.get(node_id)
+            if identity is not None and identity.node_type == "lite":
+                continue
            total_available += memory.ram_available

        return total_available
@@ -1327,6 +1397,7 @@ class API:
    async def run(self):
        shutdown_ev = anyio.Event()

+        bonjour_cleanup = self._register_bonjour_service()
        try:
            async with create_task_group() as tg:
                self._tg = tg
@@ -1342,9 +1413,48 @@ class API:
                    with anyio.CancelScope(shield=True):
                        shutdown_ev.set()
        finally:
+            bonjour_cleanup()
+            self._event_log.close()
            self.command_sender.close()
            self.global_event_receiver.close()

+    def _register_bonjour_service(self) -> Callable[[], None]:
+        """Register a Bonjour service via the system mDNSResponder. Returns a cleanup function."""
+        import subprocess
+        import sys
+
+        if sys.platform != "darwin":
+            logger.info("Bonjour service registration is only supported on macOS")
+            return lambda: None
+
+        service_name = f"EXO Cluster ({self.node_id[:8]})"
+        try:
+            proc = subprocess.Popen(
+                [
+                    "dns-sd",
+                    "-R",
+                    service_name,
+                    "_exo._tcp",
+                    "local",
+                    str(self.port),
+                    f"node_id={self.node_id}",
+                ],
+                stdout=subprocess.DEVNULL,
+                stderr=subprocess.DEVNULL,
+            )
+            logger.info(
+                f"Registered Bonjour service _exo._tcp on port {self.port} (pid {proc.pid})"
+            )
+
+            def cleanup() -> None:
+                proc.terminate()
+                proc.wait()
+
+            return cleanup
+        except Exception as e:
+            logger.warning(f"Failed to register Bonjour service: {e}")
+            return lambda: None
+
    async def run_api(self, ev: anyio.Event):
        cfg = Config()
        cfg.bind = [f"0.0.0.0:{self.port}"]
@@ -0,0 +1,187 @@
+import contextlib
+import json
+from collections import OrderedDict
+from collections.abc import Iterator
+from datetime import datetime, timezone
+from io import BufferedRandom, BufferedReader
+from pathlib import Path
+
+import msgspec
+import zstandard
+from loguru import logger
+from pydantic import TypeAdapter
+
+from exo.shared.types.events import Event
+
+_EVENT_ADAPTER: TypeAdapter[Event] = TypeAdapter(Event)
+
+_HEADER_SIZE = 4  # uint32 big-endian
+_OFFSET_CACHE_SIZE = 128
+_MAX_ARCHIVES = 5
+
+
+def _serialize_event(event: Event) -> bytes:
+    return msgspec.msgpack.encode(event.model_dump(mode="json"))
+
+
+def _deserialize_event(raw: bytes) -> Event:
+    # Decode msgpack into a Python dict, then re-encode as JSON for Pydantic.
+    # Pydantic's validate_json() uses JSON-mode coercion (e.g. string -> enum)
+    # even under strict=True, whereas validate_python() does not. Going through
+    # JSON is the only way to get correct round-trip deserialization without
+    # disabling strict mode or adding casts everywhere.
+    as_json = json.dumps(msgspec.msgpack.decode(raw, type=dict))
+    return _EVENT_ADAPTER.validate_json(as_json)
+
+
+def _unpack_header(header: bytes) -> int:
+    return int.from_bytes(header, byteorder="big")
+
+
+def _skip_record(f: BufferedReader) -> bool:
+    """Skip one length-prefixed record. Returns False on EOF."""
+    header = f.read(_HEADER_SIZE)
+    if len(header) < _HEADER_SIZE:
+        return False
+    f.seek(_unpack_header(header), 1)
+    return True
+
+
+def _read_record(f: BufferedReader) -> Event | None:
+    """Read one length-prefixed record. Returns None on EOF."""
+    header = f.read(_HEADER_SIZE)
+    if len(header) < _HEADER_SIZE:
+        return None
+    length = _unpack_header(header)
+    payload = f.read(length)
+    if len(payload) < length:
+        return None
+    return _deserialize_event(payload)
+
+
+class DiskEventLog:
+    """Append-only event log backed by a file on disk.
+
+    On-disk format: sequence of length-prefixed msgpack records.
+    Each record is [4-byte big-endian uint32 length][msgpack payload].
+
+    Uses a bounded LRU cache of event index → byte offset for efficient
+    random access without storing an offset per event.
+    """
+
+    def __init__(self, directory: Path) -> None:
+        self._directory = directory
+        self._directory.mkdir(parents=True, exist_ok=True)
+        self._active_path = directory / "events.bin"
+        self._offset_cache: OrderedDict[int, int] = OrderedDict()
+        self._count: int = 0
+
+        # Rotate stale active file from a previous session/crash
+        if self._active_path.exists():
+            self._rotate(self._active_path, self._directory)
+
+        self._file: BufferedRandom = open(self._active_path, "w+b")  # noqa: SIM115
+
+    def _cache_offset(self, idx: int, offset: int) -> None:
+        self._offset_cache[idx] = offset
+        self._offset_cache.move_to_end(idx)
+        if len(self._offset_cache) > _OFFSET_CACHE_SIZE:
+            self._offset_cache.popitem(last=False)
+
+    def _seek_to(self, f: BufferedReader, target_idx: int) -> None:
+        """Seek f to the byte offset of event target_idx, using cache or scanning forward."""
+        if target_idx in self._offset_cache:
+            self._offset_cache.move_to_end(target_idx)
+            f.seek(self._offset_cache[target_idx])
+            return
+
+        # Find the highest cached index before target_idx
+        scan_from_idx = 0
+        scan_from_offset = 0
+        for cached_idx in self._offset_cache:
+            if cached_idx < target_idx:
+                scan_from_idx = cached_idx
+                scan_from_offset = self._offset_cache[cached_idx]
+
+        # Scan forward, skipping records
+        f.seek(scan_from_offset)
+        for _ in range(scan_from_idx, target_idx):
+            _skip_record(f)
+
+        self._cache_offset(target_idx, f.tell())
+
+    def append(self, event: Event) -> None:
+        packed = _serialize_event(event)
+        self._file.write(len(packed).to_bytes(_HEADER_SIZE, byteorder="big"))
+        self._file.write(packed)
+        self._count += 1
+
+    def read_range(self, start: int, end: int) -> Iterator[Event]:
+        """Yield events from index start (inclusive) to end (exclusive)."""
+        end = min(end, self._count)
+        if start < 0 or end < 0 or start >= end:
+            return
+
+        self._file.flush()
+        with open(self._active_path, "rb") as f:
+            self._seek_to(f, start)
+            for _ in range(end - start):
+                event = _read_record(f)
+                if event is None:
+                    break
+                yield event
+
+            # Cache where we ended up so the next sequential read is a hit
+            if end < self._count:
+                self._cache_offset(end, f.tell())
+
+    def read_all(self) -> Iterator[Event]:
+        """Yield all events from the log one at a time."""
+        if self._count == 0:
+            return
+        self._file.flush()
+        with open(self._active_path, "rb") as f:
+            for _ in range(self._count):
+                event = _read_record(f)
+                if event is None:
+                    break
+                yield event
+
+    def __len__(self) -> int:
+        return self._count
+
+    def close(self) -> None:
+        """Close the file and rotate active file to compressed archive."""
+        if self._file.closed:
+            return
+        self._file.close()
+        if self._active_path.exists() and self._count > 0:
+            self._rotate(self._active_path, self._directory)
+        elif self._active_path.exists():
+            self._active_path.unlink()
+
+    @staticmethod
+    def _rotate(source: Path, directory: Path) -> None:
+        """Compress source into a timestamped archive.
+
+        Keeps at most ``_MAX_ARCHIVES`` compressed copies.  Oldest beyond
+        the limit are deleted.
+        """
+        try:
+            stamp = datetime.now(timezone.utc).strftime("%Y-%m-%d_%H-%M-%S_%f")
+            dest = directory / f"events.{stamp}.bin.zst"
+            compressor = zstandard.ZstdCompressor()
+            with open(source, "rb") as f_in, open(dest, "wb") as f_out:
+                compressor.copy_stream(f_in, f_out)
+            source.unlink()
+            logger.info(f"Rotated event log: {source} -> {dest}")
+
+            # Prune oldest archives beyond the limit
+            archives = sorted(directory.glob("events.*.bin.zst"))
+            for old in archives[:-_MAX_ARCHIVES]:
+                old.unlink()
+        except Exception as e:
+            logger.opt(exception=e).warning(f"Failed to rotate event log {source}")
+            # Clean up the source even if compression fails
+            with contextlib.suppress(OSError):
+                source.unlink()
@@ -4,6 +4,7 @@ import anyio
 from anyio.abc import TaskGroup
 from loguru import logger

+from exo.master.event_log import DiskEventLog
 from exo.master.placement import (
    add_instance_to_placements,
    cancel_unnecessary_downloads,
@@ -12,7 +13,7 @@ from exo.master.placement import (
    place_instance,
 )
 from exo.shared.apply import apply
-from exo.shared.constants import EXO_TRACING_ENABLED
+from exo.shared.constants import EXO_EVENT_LOG_DIR, EXO_TRACING_ENABLED
 from exo.shared.types.commands import (
    CreateInstance,
    DeleteInstance,
@@ -20,9 +21,11 @@ from exo.shared.types.commands import (
    ForwarderDownloadCommand,
    ImageEdits,
    ImageGeneration,
+    LiteNodeHeartbeat,
    PlaceInstance,
    RequestEventLog,
    SendInputChunk,
+    TaskCancelled,
    TaskFinished,
    TestCommand,
    TextGeneration,
@@ -38,6 +41,7 @@ from exo.shared.types.events import (
    NodeTimedOut,
    TaskCreated,
    TaskDeleted,
+    TaskStatusUpdated,
    TraceEventData,
    TracesCollected,
    TracesMerged,
@@ -88,8 +92,7 @@ class Master:
            local_event_receiver.clone_sender()
        )
        self._multi_buffer = MultiSourceBuffer[NodeId, Event]()
-        # TODO: not have this
-        self._event_log: list[Event] = []
+        self._event_log = DiskEventLog(EXO_EVENT_LOG_DIR / "master")
        self._pending_traces: dict[TaskId, dict[int, list[TraceEventData]]] = {}
        self._expected_ranks: dict[TaskId, set[int]] = {}

@@ -103,6 +106,7 @@ class Master:
                tg.start_soon(self._loopback_processor)
                tg.start_soon(self._plan)
        finally:
+            self._event_log.close()
            self.global_event_sender.close()
            self.local_event_receiver.close()
            self.command_receiver.close()
@@ -278,7 +282,7 @@ class Master:
                        case DeleteInstance():
                            placement = delete_instance(command, self.state.instances)
                            transition_events = get_transition_events(
-                                self.state.instances, placement
+                                self.state.instances, placement, self.state.tasks
                            )
                            for cmd in cancel_unnecessary_downloads(
                                placement, self.state.downloads
@@ -296,9 +300,10 @@ class Master:
                                self.state.instances,
                                self.state.node_memory,
                                self.state.node_network,
+                                node_identities=self.state.node_identities,
                            )
                            transition_events = get_transition_events(
-                                self.state.instances, placement
+                                self.state.instances, placement, self.state.tasks
                            )
                            generated_events.extend(transition_events)
                        case CreateInstance():
@@ -308,7 +313,7 @@ class Master:
                                self.state.instances,
                            )
                            transition_events = get_transition_events(
-                                self.state.instances, placement
+                                self.state.instances, placement, self.state.tasks
                            )
                            generated_events.extend(transition_events)
                        case SendInputChunk(chunk=chunk):
@@ -318,6 +323,18 @@ class Master:
                                    chunk=chunk,
                                )
                            )
+                        case TaskCancelled():
+                            if (
+                                task_id := self.command_task_mapping.get(
+                                    command.cancelled_command_id
+                                )
+                            ) is not None:
+                                generated_events.append(
+                                    TaskStatusUpdated(
+                                        task_status=TaskStatus.Cancelled,
+                                        task_id=task_id,
+                                    )
+                                )
                        case TaskFinished():
                            generated_events.append(
                                TaskDeleted(
@@ -326,20 +343,26 @@ class Master:
                                    ]
                                )
                            )
-                            if command.finished_command_id in self.command_task_mapping:
-                                del self.command_task_mapping[
-                                    command.finished_command_id
-                                ]
+                            self.command_task_mapping.pop(
+                                command.finished_command_id, None
+                            )
+                        case LiteNodeHeartbeat():
+                            generated_events.append(
+                                NodeGatheredInfo(
+                                    node_id=command.target_node_id,
+                                    when=str(datetime.now(tz=timezone.utc)),
+                                    info=command.info,
+                                )
+                            )
                        case RequestEventLog():
                            # We should just be able to send everything, since other buffers will ignore old messages
                            # rate limit to 1000 at a time
-                            for i in range(
-                                command.since_idx,
-                                min(command.since_idx + 1000, len(self._event_log)),
+                            end = min(command.since_idx + 1000, len(self._event_log))
+                            for i, event in enumerate(
+                                self._event_log.read_range(command.since_idx, end),
+                                start=command.since_idx,
                            ):
-                                await self._send_event(
-                                    IndexedEvent(idx=i, event=self._event_log[i])
-                                )
+                                await self._send_event(IndexedEvent(idx=i, event=event))
                    for event in generated_events:
                        await self.event_sender.send(event)
                except ValueError as e:
@@ -22,9 +22,15 @@ from exo.shared.types.commands import (
    PlaceInstance,
 )
 from exo.shared.types.common import NodeId
-from exo.shared.types.events import Event, InstanceCreated, InstanceDeleted
+from exo.shared.types.events import (
+    Event,
+    InstanceCreated,
+    InstanceDeleted,
+    TaskStatusUpdated,
+)
 from exo.shared.types.memory import Memory
-from exo.shared.types.profiling import MemoryUsage, NodeNetworkInfo
+from exo.shared.types.profiling import MemoryUsage, NodeIdentity, NodeNetworkInfo
+from exo.shared.types.tasks import Task, TaskId, TaskStatus
 from exo.shared.types.worker.downloads import (
    DownloadOngoing,
    DownloadProgress,
@@ -61,8 +67,21 @@ def place_instance(
    node_memory: Mapping[NodeId, MemoryUsage],
    node_network: Mapping[NodeId, NodeNetworkInfo],
    required_nodes: set[NodeId] | None = None,
+    node_identities: Mapping[NodeId, NodeIdentity] = {},
 ) -> dict[InstanceId, Instance]:
+    lite_node_ids = {
+        nid
+        for nid, identity in node_identities.items()
+        if identity.node_type == "lite"
+    }
    cycles = topology.get_cycles()
+    if lite_node_ids:
+        filtered_cycles: list[Cycle] = []
+        for cycle in cycles:
+            filtered_nodes = [nid for nid in cycle.node_ids if nid not in lite_node_ids]
+            if filtered_nodes:
+                filtered_cycles.append(Cycle(filtered_nodes))
+        cycles = filtered_cycles
    candidate_cycles = list(filter(lambda it: len(it) >= command.min_nodes, cycles))

    # Filter to cycles containing all required nodes (subset matching)
@@ -186,6 +205,7 @@ def delete_instance(
 def get_transition_events(
    current_instances: Mapping[InstanceId, Instance],
    target_instances: Mapping[InstanceId, Instance],
+    tasks: Mapping[TaskId, Task],
 ) -> Sequence[Event]:
    events: list[Event] = []

@@ -201,6 +221,18 @@ def get_transition_events(
    # find instances to delete
    for instance_id in current_instances:
        if instance_id not in target_instances:
+            for task in tasks.values():
+                if task.instance_id == instance_id and task.task_status in [
+                    TaskStatus.Pending,
+                    TaskStatus.Running,
+                ]:
+                    events.append(
+                        TaskStatusUpdated(
+                            task_status=TaskStatus.Cancelled,
+                            task_id=task.task_id,
+                        )
+                    )
+
            events.append(
                InstanceDeleted(
                    instance_id=instance_id,
@@ -4,7 +4,11 @@ import json
 from collections.abc import AsyncGenerator
 from typing import Any, cast

-from exo.master.adapters.claude import collect_claude_response, generate_claude_stream
+from exo.master.adapters.claude import (
+    ClaudeMessagesResponse,
+    collect_claude_response,
+    generate_claude_stream,
+)
 from exo.shared.types.api import ToolCallItem
 from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
 from exo.shared.types.common import CommandId, ModelId
@@ -17,6 +21,18 @@ async def _chunks_to_stream(
        yield chunk


+async def _collect_response(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> ClaudeMessagesResponse:
+    """Helper to consume the async generator and parse the JSON response."""
+    parts: list[str] = []
+    async for part in collect_claude_response(command_id, model, chunk_stream):
+        parts.append(part)
+    return ClaudeMessagesResponse.model_validate_json("".join(parts))
+
+
 MODEL = ModelId("test-model")
 COMMAND_ID = CommandId("cmd_test123")

@@ -47,7 +63,7 @@ class TestCollectClaudeResponseToolUse:
                ],
            ),
        ]
-        response = await collect_claude_response(
+        response = await _collect_response(
            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
        )

@@ -77,7 +93,7 @@ class TestCollectClaudeResponseToolUse:
                ],
            ),
        ]
-        response = await collect_claude_response(
+        response = await _collect_response(
            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
        )

@@ -102,7 +118,7 @@ class TestCollectClaudeResponseToolUse:
                ],
            ),
        ]
-        response = await collect_claude_response(
+        response = await _collect_response(
            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
        )

@@ -116,7 +132,7 @@ class TestCollectClaudeResponseToolUse:

    async def test_no_content_produces_empty_text_block(self):
        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = []
-        response = await collect_claude_response(
+        response = await _collect_response(
            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
        )
        assert len(response.content) == 1
@@ -0,0 +1,155 @@
+from pathlib import Path
+
+import pytest
+
+from exo.master.event_log import DiskEventLog
+from exo.shared.types.events import TestEvent
+
+
+@pytest.fixture
+def log_dir(tmp_path: Path) -> Path:
+    return tmp_path / "event_log"
+
+
+def test_append_and_read_back(log_dir: Path):
+    log = DiskEventLog(log_dir)
+    events = [TestEvent() for _ in range(5)]
+    for e in events:
+        log.append(e)
+
+    assert len(log) == 5
+
+    result = list(log.read_all())
+    assert len(result) == 5
+    for original, restored in zip(events, result, strict=True):
+        assert original.event_id == restored.event_id
+
+    log.close()
+
+
+def test_read_range(log_dir: Path):
+    log = DiskEventLog(log_dir)
+    events = [TestEvent() for _ in range(10)]
+    for e in events:
+        log.append(e)
+
+    result = list(log.read_range(3, 7))
+    assert len(result) == 4
+    for i, restored in enumerate(result):
+        assert events[3 + i].event_id == restored.event_id
+
+    log.close()
+
+
+def test_read_range_bounds(log_dir: Path):
+    log = DiskEventLog(log_dir)
+    events = [TestEvent() for _ in range(3)]
+    for e in events:
+        log.append(e)
+
+    # Start beyond count
+    assert list(log.read_range(5, 10)) == []
+    # Negative start
+    assert list(log.read_range(-1, 2)) == []
+    # End beyond count is clamped
+    result = list(log.read_range(1, 100))
+    assert len(result) == 2
+
+    log.close()
+
+
+def test_empty_log(log_dir: Path):
+    log = DiskEventLog(log_dir)
+    assert len(log) == 0
+    assert list(log.read_all()) == []
+    assert list(log.read_range(0, 10)) == []
+    log.close()
+
+
+def _archives(log_dir: Path) -> list[Path]:
+    return sorted(log_dir.glob("events.*.bin.zst"))
+
+
+def test_rotation_on_close(log_dir: Path):
+    log = DiskEventLog(log_dir)
+    log.append(TestEvent())
+    log.close()
+
+    active = log_dir / "events.bin"
+    assert not active.exists()
+
+    archives = _archives(log_dir)
+    assert len(archives) == 1
+    assert archives[0].stat().st_size > 0
+
+
+def test_rotation_on_construction_with_stale_file(log_dir: Path):
+    log_dir.mkdir(parents=True, exist_ok=True)
+    (log_dir / "events.bin").write_bytes(b"stale data")
+
+    log = DiskEventLog(log_dir)
+    archives = _archives(log_dir)
+    assert len(archives) == 1
+    assert archives[0].exists()
+    assert len(log) == 0
+
+    log.close()
+
+
+def test_empty_log_no_archive(log_dir: Path):
+    """Closing an empty log should not leave an archive."""
+    log = DiskEventLog(log_dir)
+    log.close()
+
+    active = log_dir / "events.bin"
+
+    assert not active.exists()
+    assert _archives(log_dir) == []
+
+
+def test_close_is_idempotent(log_dir: Path):
+    log = DiskEventLog(log_dir)
+    log.append(TestEvent())
+    log.close()
+    archive = _archives(log_dir)
+    log.close()  # should not raise
+
+    assert _archives(log_dir) == archive
+
+
+def test_successive_sessions(log_dir: Path):
+    """Simulate two master sessions: both archives should be kept."""
+    log1 = DiskEventLog(log_dir)
+    log1.append(TestEvent())
+    log1.close()
+
+    first_archive = _archives(log_dir)[-1]
+
+    log2 = DiskEventLog(log_dir)
+    log2.append(TestEvent())
+    log2.append(TestEvent())
+    log2.close()
+
+    # Session 1 archive shifted to slot 2, session 2 in slot 1
+    second_archive = _archives(log_dir)[-1]
+    should_be_first_archive = _archives(log_dir)[-2]
+
+    assert first_archive.exists()
+    assert second_archive.exists()
+    assert first_archive != second_archive
+    assert should_be_first_archive == first_archive
+
+
+def test_rotation_keeps_at_most_5_archives(log_dir: Path):
+    """After 7 sessions, only the 5 most recent archives should remain."""
+    all_archives: list[Path] = []
+    for _ in range(7):
+        log = DiskEventLog(log_dir)
+        log.append(TestEvent())
+        log.close()
+        all_archives.append(_archives(log_dir)[-1])
+
+    for old in all_archives[:2]:
+        assert not old.exists()
+    for recent in all_archives[2:]:
+        assert recent.exists()
@@ -239,7 +239,7 @@ def test_get_transition_events_no_change(instance: Instance):
    target_instances = {instance_id: instance}

    # act
-    events = get_transition_events(current_instances, target_instances)
+    events = get_transition_events(current_instances, target_instances, {})

    # assert
    assert len(events) == 0
@@ -252,7 +252,7 @@ def test_get_transition_events_create_instance(instance: Instance):
    target_instances: dict[InstanceId, Instance] = {instance_id: instance}

    # act
-    events = get_transition_events(current_instances, target_instances)
+    events = get_transition_events(current_instances, target_instances, {})

    # assert
    assert len(events) == 1
@@ -266,7 +266,7 @@ def test_get_transition_events_delete_instance(instance: Instance):
    target_instances: dict[InstanceId, Instance] = {}

    # act
-    events = get_transition_events(current_instances, target_instances)
+    events = get_transition_events(current_instances, target_instances, {})

    # assert
    assert len(events) == 1
@@ -207,10 +207,10 @@ class Router:
                try:
                    logger.trace(f"Sending message on {topic} with payload {data}")
                    await self._net.gossipsub_publish(topic, data)
-                # As a hack, this also catches AllQueuesFull
-                # Need to fix that ASAP.
-                except (NoPeersSubscribedToTopicError, AllQueuesFullError):
+                except NoPeersSubscribedToTopicError:
                    pass
+                except AllQueuesFullError:
+                    logger.warning(f"All peer queues full, dropping message on {topic}")


 def get_node_id_keypair(
@@ -31,6 +31,7 @@ from exo.shared.types.events import (
 from exo.shared.types.profiling import (
    NodeIdentity,
    NodeNetworkInfo,
+    NodeRdmaCtlStatus,
    NodeThunderboltInfo,
    ThunderboltBridgeStatus,
 )
@@ -41,13 +42,16 @@ from exo.shared.types.worker.downloads import DownloadProgress
 from exo.shared.types.worker.instances import Instance, InstanceId
 from exo.shared.types.worker.runners import RunnerId, RunnerStatus
 from exo.utils.info_gatherer.info_gatherer import (
+    LiteNodeRegistration,
    MacmonMetrics,
    MacThunderboltConnections,
    MacThunderboltIdentifiers,
    MemoryUsage,
    MiscData,
    NodeConfig,
+    NodeDiskUsage,
    NodeNetworkInterfaces,
+    RdmaCtlStatus,
    StaticNodeInformation,
    ThunderboltBridgeInfo,
 )
@@ -215,14 +219,12 @@ def apply_node_timed_out(event: NodeTimedOut, state: State) -> State:
        key: value for key, value in state.downloads.items() if key != event.node_id
    }
    # Clean up all granular node mappings
-    node_identities = {
-        key: value
-        for key, value in state.node_identities.items()
-        if key != event.node_id
-    }
    node_memory = {
        key: value for key, value in state.node_memory.items() if key != event.node_id
    }
+    node_disk = {
+        key: value for key, value in state.node_disk.items() if key != event.node_id
+    }
    node_system = {
        key: value for key, value in state.node_system.items() if key != event.node_id
    }
@@ -239,6 +241,14 @@ def apply_node_timed_out(event: NodeTimedOut, state: State) -> State:
        for key, value in state.node_thunderbolt_bridge.items()
        if key != event.node_id
    }
+    node_rdma_ctl = {
+        key: value for key, value in state.node_rdma_ctl.items() if key != event.node_id
+    }
+    node_identities = {
+        key: value
+        for key, value in state.node_identities.items()
+        if key != event.node_id
+    }
    # Only recompute cycles if the leaving node had TB bridge enabled
    leaving_node_status = state.node_thunderbolt_bridge.get(event.node_id)
    leaving_node_had_tb_enabled = (
@@ -254,12 +264,14 @@ def apply_node_timed_out(event: NodeTimedOut, state: State) -> State:
            "downloads": downloads,
            "topology": topology,
            "last_seen": last_seen,
-            "node_identities": node_identities,
            "node_memory": node_memory,
+            "node_disk": node_disk,
            "node_system": node_system,
            "node_network": node_network,
            "node_thunderbolt": node_thunderbolt,
            "node_thunderbolt_bridge": node_thunderbolt_bridge,
+            "node_rdma_ctl": node_rdma_ctl,
+            "node_identities": node_identities,
            "thunderbolt_bridge_cycles": thunderbolt_bridge_cycles,
        }
    )
@@ -288,6 +300,8 @@ def apply_node_gathered_info(event: NodeGatheredInfo, state: State) -> State:
            update["node_memory"] = {**state.node_memory, event.node_id: info.memory}
        case MemoryUsage():
            update["node_memory"] = {**state.node_memory, event.node_id: info}
+        case NodeDiskUsage():
+            update["node_disk"] = {**state.node_disk, event.node_id: info.disk_usage}
        case NodeConfig():
            pass
        case MiscData():
@@ -302,7 +316,12 @@ def apply_node_gathered_info(event: NodeGatheredInfo, state: State) -> State:
        case StaticNodeInformation():
            current_identity = state.node_identities.get(event.node_id, NodeIdentity())
            new_identity = current_identity.model_copy(
-                update={"model_id": info.model, "chip_id": info.chip}
+                update={
+                    "model_id": info.model,
+                    "chip_id": info.chip,
+                    "os_version": info.os_version,
+                    "os_build_version": info.os_build_version,
+                }
            )
            update["node_identities"] = {
                **state.node_identities,
@@ -354,6 +373,35 @@ def apply_node_gathered_info(event: NodeGatheredInfo, state: State) -> State:
                        new_tb_bridge, state.node_network
                    )
                )
+        case RdmaCtlStatus():
+            update["node_rdma_ctl"] = {
+                **state.node_rdma_ctl,
+                event.node_id: NodeRdmaCtlStatus(enabled=info.enabled),
+            }
+        case LiteNodeRegistration():
+            current_identity = state.node_identities.get(event.node_id, NodeIdentity())
+            new_identity = current_identity.model_copy(
+                update={
+                    "model_id": info.model,
+                    "chip_id": info.chip,
+                    "os_version": info.os_version,
+                    "friendly_name": info.friendly_name,
+                    "node_type": "lite",
+                }
+            )
+            update["node_identities"] = {
+                **state.node_identities,
+                event.node_id: new_identity,
+            }
+            update["node_memory"] = {
+                **state.node_memory,
+                event.node_id: MemoryUsage.from_bytes(
+                    ram_total=info.ram_total,
+                    ram_available=info.ram_available,
+                    swap_total=0,
+                    swap_available=0,
+                ),
+            }

    return state.model_copy(update=update)

@@ -43,7 +43,8 @@ DASHBOARD_DIR = (
 )

 # Log files (data/logs or cache)
-EXO_LOG = EXO_CACHE_HOME / "exo.log"
+EXO_LOG_DIR = EXO_CACHE_HOME / "exo_log"
+EXO_LOG = EXO_LOG_DIR / "exo.log"
 EXO_TEST_LOG = EXO_CACHE_HOME / "exo_test.log"

 # Identity (config)
@@ -60,6 +61,7 @@ EXO_MAX_CHUNK_SIZE = 512 * 1024

 EXO_CUSTOM_MODEL_CARDS_DIR = EXO_DATA_HOME / "custom_model_cards"

+EXO_EVENT_LOG_DIR = EXO_DATA_HOME / "event_log"
 EXO_IMAGE_CACHE_DIR = EXO_CACHE_HOME / "images"
 EXO_TRACING_CACHE_DIR = EXO_CACHE_HOME / "traces"

@@ -1,11 +1,30 @@
 import logging
 import sys
+from collections.abc import Iterator
 from pathlib import Path

+import zstandard
 from hypercorn import Config
 from hypercorn.logging import Logger as HypercornLogger
 from loguru import logger

+_MAX_LOG_ARCHIVES = 5
+
+
+def _zstd_compress(filepath: str) -> None:
+    source = Path(filepath)
+    dest = source.with_suffix(source.suffix + ".zst")
+    cctx = zstandard.ZstdCompressor()
+    with open(source, "rb") as f_in, open(dest, "wb") as f_out:
+        cctx.copy_stream(f_in, f_out)
+    source.unlink()
+
+
+def _once_then_never() -> Iterator[bool]:
+    yield True
+    while True:
+        yield False
+

 class InterceptLogger(HypercornLogger):
    def __init__(self, config: Config):
@@ -53,13 +72,16 @@ def logger_setup(log_file: Path | None, verbosity: int = 0):
            enqueue=True,
        )
    if log_file:
+        rotate_once = _once_then_never()
        logger.add(
            log_file,
            format="[ {time:YYYY-MM-DD HH:mm:ss.SSS} | {level: <8} | {name}:{function}:{line} ] {message}",
            level="INFO",
            colorize=False,
            enqueue=True,
-            rotation="1 week",
+            rotation=lambda _, __: next(rotate_once),
+            retention=_MAX_LOG_ARCHIVES,
+            compression=_zstd_compress,
        )


@@ -30,11 +30,11 @@ from exo.utils.pydantic_ext import CamelCaseModel
 # kinda ugly...
 # TODO: load search path from config.toml
 _custom_cards_dir = Path(str(EXO_CUSTOM_MODEL_CARDS_DIR))
-_csp = [Path(RESOURCES_DIR) / "inference_model_cards", _custom_cards_dir]
-if EXO_ENABLE_IMAGE_MODELS:
-    _csp.append(Path(RESOURCES_DIR) / "image_model_cards")
-
-CARD_SEARCH_PATH = _csp
+CARD_SEARCH_PATH = [
+    Path(RESOURCES_DIR) / "inference_model_cards",
+    Path(RESOURCES_DIR) / "image_model_cards",
+    _custom_cards_dir,
+]

 _card_cache: dict[ModelId, "ModelCard"] = {}

@@ -49,10 +49,16 @@ async def _refresh_card_cache():
                pass


+def _is_image_card(card: "ModelCard") -> bool:
+    return any(t in (ModelTask.TextToImage, ModelTask.ImageToImage) for t in card.tasks)
+
+
 async def get_model_cards() -> list["ModelCard"]:
    if len(_card_cache) == 0:
        await _refresh_card_cache()
-    return list(_card_cache.values())
+    if EXO_ENABLE_IMAGE_MODELS:
+        return list(_card_cache.values())
+    return [c for c in _card_cache.values() if not _is_image_card(c)]


 class ModelTask(str, Enum):
@@ -183,6 +189,7 @@ class ConfigData(BaseModel):
            ["MiniMaxM2ForCausalLM"],
            ["LlamaForCausalLM"],
            ["GptOssForCausalLM"],
+            ["Step3p5ForCausalLM"],
        ]

    @model_validator(mode="before")
@@ -3,8 +3,7 @@ from collections.abc import Generator
 from typing import Annotated, Any, Literal
 from uuid import uuid4

-from pydantic import BaseModel, Field, field_validator
-from pydantic_core import PydanticUseDefault
+from pydantic import BaseModel, Field

 from exo.shared.models.model_cards import ModelCard, ModelId
 from exo.shared.types.common import CommandId, NodeId
@@ -199,6 +198,7 @@ class ChatCompletionRequest(BaseModel):
    top_p: float | None = None
    top_k: int | None = None
    tools: list[dict[str, Any]] | None = None
+    enable_thinking: bool | None = None
    tool_choice: str | dict[str, Any] | None = None
    parallel_tool_calls: bool | None = None
    user: str | None = None
@@ -227,13 +227,6 @@ class PlaceInstanceParams(BaseModel):
    instance_meta: InstanceMeta = InstanceMeta.MlxRing
    min_nodes: int = 1

-    @field_validator("sharding", "instance_meta", mode="plain")
-    @classmethod
-    def use_default(cls, v: object):
-        if not v or not isinstance(v, (Sharding, InstanceMeta)):
-            raise PydanticUseDefault()
-        return v
-

 class CreateInstanceParams(BaseModel):
    instance: Instance
@@ -413,3 +406,13 @@ class TraceListItem(CamelCaseModel):

 class TraceListResponse(CamelCaseModel):
    traces: list[TraceListItem]
+
+
+class LiteNodeHeartbeatRequest(CamelCaseModel):
+    node_id: str
+    model: str
+    chip: str
+    os_version: str
+    friendly_name: str
+    ram_total: int
+    ram_available: int
@@ -10,6 +10,7 @@ from exo.shared.types.common import CommandId, NodeId
 from exo.shared.types.text_generation import TextGenerationTaskParams
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding, ShardMetadata
+from exo.utils.info_gatherer.info_gatherer import GatheredInfo
 from exo.utils.pydantic_ext import CamelCaseModel, TaggedModel


@@ -48,6 +49,10 @@ class DeleteInstance(BaseCommand):
    instance_id: InstanceId


+class TaskCancelled(BaseCommand):
+    cancelled_command_id: CommandId
+
+
 class TaskFinished(BaseCommand):
    finished_command_id: CommandId

@@ -58,6 +63,11 @@ class SendInputChunk(BaseCommand):
    chunk: InputImageChunk


+class LiteNodeHeartbeat(BaseCommand):
+    target_node_id: NodeId
+    info: GatheredInfo
+
+
 class RequestEventLog(BaseCommand):
    since_idx: int

@@ -89,8 +99,10 @@ Command = (
    | PlaceInstance
    | CreateInstance
    | DeleteInstance
+    | TaskCancelled
    | TaskFinished
    | SendInputChunk
+    | LiteNodeHeartbeat
 )


@@ -1,4 +1,6 @@
+import shutil
 from collections.abc import Sequence
+from pathlib import Path
 from typing import Literal, Self

 import psutil
@@ -38,6 +40,22 @@ class MemoryUsage(CamelCaseModel):
        )


+class DiskUsage(CamelCaseModel):
+    """Disk space usage for the models directory."""
+
+    total: Memory
+    available: Memory
+
+    @classmethod
+    def from_path(cls, path: Path) -> Self:
+        """Get disk usage stats for the partition containing path."""
+        total, _used, free = shutil.disk_usage(path)
+        return cls(
+            total=Memory.from_bytes(total),
+            available=Memory.from_bytes(free),
+        )
+
+
 class SystemPerformanceProfile(CamelCaseModel):
    # TODO: flops_fp16: float

@@ -50,6 +68,8 @@ class SystemPerformanceProfile(CamelCaseModel):

 InterfaceType = Literal["wifi", "ethernet", "maybe_ethernet", "thunderbolt", "unknown"]

+NodeType = Literal["full", "lite"]
+

 class NetworkInterfaceInfo(CamelCaseModel):
    name: str
@@ -63,6 +83,9 @@ class NodeIdentity(CamelCaseModel):
    model_id: str = "Unknown"
    chip_id: str = "Unknown"
    friendly_name: str = "Unknown"
+    os_version: str = "Unknown"
+    os_build_version: str = "Unknown"
+    node_type: NodeType = "full"


 class NodeNetworkInfo(CamelCaseModel):
@@ -77,6 +100,12 @@ class NodeThunderboltInfo(CamelCaseModel):
    interfaces: Sequence[ThunderboltIdentifier] = []


+class NodeRdmaCtlStatus(CamelCaseModel):
+    """Whether RDMA is enabled on this node (via rdma_ctl)."""
+
+    enabled: bool
+
+
 class ThunderboltBridgeStatus(CamelCaseModel):
    """Whether the Thunderbolt Bridge network service is enabled on this node."""

@@ -8,9 +8,11 @@ from pydantic.alias_generators import to_camel
 from exo.shared.topology import Topology, TopologySnapshot
 from exo.shared.types.common import NodeId
 from exo.shared.types.profiling import (
+    DiskUsage,
    MemoryUsage,
    NodeIdentity,
    NodeNetworkInfo,
+    NodeRdmaCtlStatus,
    NodeThunderboltInfo,
    SystemPerformanceProfile,
    ThunderboltBridgeStatus,
@@ -49,10 +51,12 @@ class State(CamelCaseModel):
    # Granular node state mappings (update independently at different frequencies)
    node_identities: Mapping[NodeId, NodeIdentity] = {}
    node_memory: Mapping[NodeId, MemoryUsage] = {}
+    node_disk: Mapping[NodeId, DiskUsage] = {}
    node_system: Mapping[NodeId, SystemPerformanceProfile] = {}
    node_network: Mapping[NodeId, NodeNetworkInfo] = {}
    node_thunderbolt: Mapping[NodeId, NodeThunderboltInfo] = {}
    node_thunderbolt_bridge: Mapping[NodeId, ThunderboltBridgeStatus] = {}
+    node_rdma_ctl: Mapping[NodeId, NodeRdmaCtlStatus] = {}

    # Detected cycles where all nodes have Thunderbolt bridge enabled (>2 nodes)
    thunderbolt_bridge_cycles: Sequence[Sequence[NodeId]] = []
@@ -24,6 +24,7 @@ class TaskStatus(str, Enum):
    Complete = "Complete"
    TimedOut = "TimedOut"
    Failed = "Failed"
+    Cancelled = "Cancelled"


 class BaseTask(TaggedModel):
@@ -60,6 +61,11 @@ class TextGeneration(BaseTask):  # emitted by Master
    error_message: str | None = Field(default=None)


+class CancelTask(BaseTask):
+    cancelled_task_id: TaskId
+    runner_id: RunnerId
+
+
 class ImageGeneration(BaseTask):  # emitted by Master
    command_id: CommandId
    task_params: ImageGenerationTaskParams
@@ -87,6 +93,7 @@ Task = (
    | LoadModel
    | StartWarmup
    | TextGeneration
+    | CancelTask
    | ImageGeneration
    | ImageEdits
    | Shutdown
@@ -40,5 +40,6 @@ class TextGenerationTaskParams(BaseModel, frozen=True):
    stop: str | list[str] | None = None
    seed: int | None = None
    chat_template_messages: list[dict[str, Any]] | None = None
+    enable_thinking: bool | None = None
    logprobs: bool = False
    top_logprobs: int | None = None
@@ -12,6 +12,7 @@ class ThunderboltConnection(CamelCaseModel):
 class ThunderboltIdentifier(CamelCaseModel):
    rdma_interface: str
    domain_uuid: str
+    link_speed: str = ""


 ## Intentionally minimal, only collecting data we care about - there's a lot more
@@ -19,6 +20,7 @@ class ThunderboltIdentifier(CamelCaseModel):

 class _ReceptacleTag(BaseModel, extra="ignore"):
    receptacle_id_key: str | None = None
+    current_speed_key: str | None = None


 class _ConnectivityItem(BaseModel, extra="ignore"):
@@ -42,7 +44,9 @@ class ThunderboltConnectivityData(BaseModel, extra="ignore"):
        # if tag not in ifaces: return None
        iface = f"rdma_{ifaces[tag]}"
        return ThunderboltIdentifier(
-            rdma_interface=iface, domain_uuid=self.domain_uuid_key
+            rdma_interface=iface,
+            domain_uuid=self.domain_uuid_key,
+            link_speed=self.receptacle_1_tag.current_speed_key or "",
        )

    def conn(self) -> ThunderboltConnection | None:
@@ -26,6 +26,7 @@ class DownloadProgressData(CamelCaseModel):
 class BaseDownloadProgress(TaggedModel):
    node_id: NodeId
    shard_metadata: ShardMetadata
+    model_directory: str = ""


 class DownloadPending(BaseDownloadProgress):
@@ -62,6 +62,7 @@ class PartialImageResponse(BaseRunnerResponse):
 class ToolCallResponse(BaseRunnerResponse):
    tool_calls: list[ToolCallItem]
    usage: Usage | None
+    stats: GenerationStats | None = None


 class FinishedResponse(BaseRunnerResponse):
@@ -1,5 +1,7 @@
+import sys
+
+
 def print_startup_banner(port: int) -> None:
-    """Print a prominent startup banner with API endpoint information."""
    dashboard_url = f"http://localhost:{port}"
    banner = f"""
 ╔═══════════════════════════════════════════════════════════════════════╗
@@ -27,4 +29,4 @@ def print_startup_banner(port: int) -> None:

 """

-    print(banner)
+    print(banner, file=sys.stderr)
--- a/Show More
+++ b/Show More