feat: add MetaInstance declarative layer with reconciliation

Adds a declarative MetaInstance system for managing model instances with
automatic placement, retry logic (max 3 attempts), and lifecycle
management via a reconciliation loop.

- Process managers for instance health, meta-instance lifecycle, and
  node timeout detection
- Reconciliation engine driving state transitions and cascading deletes
- Dashboard UI for creating/managing MetaInstances with node selection,
  sharding config, retry status, and error feedback
- JACCL SideChannel integration for distributed inference coordination
- Comprehensive test suite (25+ edge cases)

Split from original #1519. Independent bug fixes extracted to:
  #1547 (misc fixes), #1546 (JACCL sidechannel), #1582 (download
  detection), #1580 (RDMA warning)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

.github/workflows/build-app.yml

@@ -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"

.github/workflows/pipeline.yml

@@ -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 }}

.mlx_typings/mlx/core/__init__.pyi

@@ -2366,7 +2366,7 @@ class custom_function:
 def default_device() -> Device:
     """Get the default device."""
 
-def default_stream(device: Device) -> Stream:
+def default_stream(device: Device | DeviceType) -> Stream:
     """Get the device's default stream."""
 
 def degrees(a: array, /, *, stream: Stream | Device | None = ...) -> array:

.mlx_typings/mlx/nn/layers/base.pyi

@@ -200,7 +200,7 @@ class Module(dict):
     ) -> mx.MX_ARRAY_TREE:  # -> dict[Any, Any | dict[Any, Any | dict[Any, Any] | list[Any]] | dict[Any, Any] | list[Any]]:
         """Return the submodules that do not contain other modules."""
 
-    def update(self, parameters: dict, strict: bool = ...) -> Module:
+    def update(self, parameters: dict[str, Any], strict: bool = ...) -> Module:
         """Replace the parameters of this Module with the provided ones in the
         dict of dicts and lists.
 

.mlx_typings/mlx/utils.pyi

@@ -7,7 +7,10 @@ from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 from mlx.core import MX_ARRAY_TREE
 
 def tree_map(
-    fn: Callable, tree: Any, *rest: Any, is_leaf: Optional[Callable] = ...
+    fn: Callable[..., Any],
+    tree: Any,
+    *rest: Any,
+    is_leaf: Callable[..., bool] | None = ...,
 ) -> Any:
     """Applies ``fn`` to the leaves of the Python tree ``tree`` and
     returns a new collection with the results.
@@ -44,11 +47,11 @@ def tree_map(
     """
 
 def tree_map_with_path(
-    fn: Callable,
+    fn: Callable[..., Any],
     tree: Any,
     *rest: Any,
-    is_leaf: Optional[Callable] = ...,
-    path: Optional[Any] = ...,
+    is_leaf: Callable[..., bool] | None = ...,
+    path: str | None = ...,
 ) -> Any:
     """Applies ``fn`` to the path and leaves of the Python tree ``tree`` and
     returns a new collection with the results.
@@ -80,9 +83,9 @@ def tree_map_with_path(
 def tree_flatten(
     tree: Any,
     prefix: str = ...,
-    is_leaf: Optional[Callable] = ...,
-    destination: Optional[Union[List[Tuple[str, Any]], Dict[str, Any]]] = ...,
-) -> Union[List[Tuple[str, Any]], Dict[str, Any]]:
+    is_leaf: Callable[..., bool] | None = ...,
+    destination: list[tuple[str, Any]] | dict[str, Any] | None = ...,
+) -> list[tuple[str, Any]] | dict[str, Any]:
     """Flattens a Python tree to a list of key, value tuples.
 
     The keys are using the dot notation to define trees of arbitrary depth and
@@ -118,7 +121,7 @@ def tree_flatten(
             the Python tree.
     """
 
-def tree_unflatten(tree: Union[List[Tuple[str, Any]], Dict[str, Any]]) -> Any:
+def tree_unflatten(tree: list[tuple[str, Any]] | dict[str, Any]) -> Any:
     """Recreate a Python tree from its flat representation.
 
     .. code-block:: python

.mlx_typings/mlx_lm/models/glm_moe_dsa.pyi

@@ -0,0 +1,46 @@
+"""Type stubs for mlx_lm.models.glm_moe_dsa"""
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+from .base import BaseModelArgs
+from .deepseek_v32 import Model as DSV32Model
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    vocab_size: int
+    hidden_size: int
+    index_head_dim: int
+    index_n_heads: int
+    index_topk: int
+    intermediate_size: int
+    moe_intermediate_size: int
+    num_hidden_layers: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    n_shared_experts: Optional[int]
+    n_routed_experts: Optional[int]
+    routed_scaling_factor: float
+    kv_lora_rank: int
+    q_lora_rank: int
+    qk_rope_head_dim: int
+    v_head_dim: int
+    qk_nope_head_dim: int
+    topk_method: str
+    scoring_func: str
+    norm_topk_prob: bool
+    n_group: int
+    topk_group: int
+    num_experts_per_tok: int
+    moe_layer_freq: int
+    first_k_dense_replace: int
+    max_position_embeddings: int
+    rms_norm_eps: float
+    rope_parameters: Dict[str, Any]
+    attention_bias: bool
+    rope_scaling: Dict[str, Any] | None
+    rope_theta: float | None
+
+class Model(DSV32Model):
+    def __init__(self, config: ModelArgs) -> None: ...

.mlx_typings/mlx_lm/models/step3p5.pyi

@@ -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: ...

AGENTS.md

@@ -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.

Cargo.lock

@@ -141,12 +141,6 @@ version = "0.3.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "76a2e8124351fda1ef8aaaa3bbd7ebbcb486bbcd4225aca0aa0d84bb2db8fecb"
 
-[[package]]
-name = "arrayvec"
-version = "0.7.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7c02d123df017efcdfbd739ef81735b36c5ba83ec3c59c80a9d7ecc718f92e50"
-
 [[package]]
 name = "asn1-rs"
 version = "0.7.1"
@@ -304,19 +298,6 @@ version = "1.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "55248b47b0caf0546f7988906588779981c43bb1bc9d0c44087278f80cdb44ba"
 
-[[package]]
-name = "bigdecimal"
-version = "0.4.9"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "560f42649de9fa436b73517378a147ec21f6c997a546581df4b4b31677828934"
-dependencies = [
- "autocfg",
- "libm",
- "num-bigint",
- "num-integer",
- "num-traits",
-]
-
 [[package]]
 name = "bimap"
 version = "0.6.3"
@@ -353,31 +334,6 @@ 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"
@@ -541,15 +497,6 @@ version = "0.4.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2f421161cb492475f1661ddc9815a745a1c894592070661180fdec3d4872e9c3"
 
-[[package]]
-name = "convert_case"
-version = "0.10.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "633458d4ef8c78b72454de2d54fd6ab2e60f9e02be22f3c6104cdc8a4e0fceb9"
-dependencies = [
- "unicode-segmentation",
-]
-
 [[package]]
 name = "core-foundation"
 version = "0.9.4"
@@ -700,41 +647,6 @@ 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"
@@ -806,29 +718,6 @@ dependencies = [
  "powerfmt",
 ]
 
-[[package]]
-name = "derive_more"
-version = "2.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "10b768e943bed7bf2cab53df09f4bc34bfd217cdb57d971e769874c9a6710618"
-dependencies = [
- "derive_more-impl",
-]
-
-[[package]]
-name = "derive_more-impl"
-version = "2.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6d286bfdaf75e988b4a78e013ecd79c581e06399ab53fbacd2d916c2f904f30b"
-dependencies = [
- "convert_case",
- "proc-macro2",
- "quote",
- "rustc_version",
- "syn 2.0.111",
- "unicode-xid",
-]
-
 [[package]]
 name = "digest"
 version = "0.10.7"
@@ -999,22 +888,17 @@ name = "exo_pyo3_bindings"
 version = "0.0.1"
 dependencies = [
  "delegate",
- "derive_more",
  "env_logger",
  "extend",
- "futures",
- "impl-trait-for-tuples",
+ "futures-lite",
  "libp2p",
  "log",
  "networking",
- "once_cell",
  "pin-project",
  "pyo3",
  "pyo3-async-runtimes",
  "pyo3-log",
  "pyo3-stub-gen",
- "thiserror 2.0.17",
- "thread_local",
  "tokio",
  "util",
 ]
@@ -1030,6 +914,12 @@ dependencies = [
  "syn 2.0.111",
 ]
 
+[[package]]
+name = "fastrand"
+version = "2.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "37909eebbb50d72f9059c3b6d82c0463f2ff062c9e95845c43a6c9c0355411be"
+
 [[package]]
 name = "ff"
 version = "0.13.1"
@@ -1138,7 +1028,10 @@ version = "2.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f78e10609fe0e0b3f4157ffab1876319b5b0db102a2c60dc4626306dc46b44ad"
 dependencies = [
+ "fastrand",
  "futures-core",
+ "futures-io",
+ "parking",
  "pin-project-lite",
 ]
 
@@ -1625,12 +1518,6 @@ 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"
@@ -1706,17 +1593,6 @@ dependencies = [
  "xmltree",
 ]
 
-[[package]]
-name = "impl-trait-for-tuples"
-version = "0.2.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a0eb5a3343abf848c0984fe4604b2b105da9539376e24fc0a3b0007411ae4fd9"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn 2.0.111",
-]
-
 [[package]]
 name = "indexmap"
 version = "2.12.1"
@@ -1745,15 +1621,6 @@ 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"
@@ -1904,12 +1771,6 @@ version = "0.2.178"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "37c93d8daa9d8a012fd8ab92f088405fb202ea0b6ab73ee2482ae66af4f42091"
 
-[[package]]
-name = "libm"
-version = "0.2.15"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f9fbbcab51052fe104eb5e5d351cf728d30a5be1fe14d9be8a3b097481fb97de"
-
 [[package]]
 name = "libp2p"
 version = "0.56.0"
@@ -2899,16 +2760,13 @@ name = "networking"
 version = "0.0.1"
 dependencies = [
  "delegate",
- "derive_more",
  "either",
  "extend",
- "futures",
+ "futures-lite",
  "futures-timer",
- "impl-trait-for-tuples",
  "keccak-const",
  "libp2p",
  "log",
- "thiserror 2.0.17",
  "tokio",
  "tracing-subscriber",
  "util",
@@ -2993,17 +2851,6 @@ dependencies = [
  "num-traits",
 ]
 
-[[package]]
-name = "num-rational"
-version = "0.4.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f83d14da390562dca69fc84082e73e548e1ad308d24accdedd2720017cb37824"
-dependencies = [
- "num-bigint",
- "num-integer",
- "num-traits",
-]
-
 [[package]]
 name = "num-traits"
 version = "0.2.19"
@@ -3307,16 +3154,6 @@ 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"
@@ -3364,28 +3201,14 @@ version = "0.27.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ab53c047fcd1a1d2a8820fe84f05d6be69e9526be40cb03b73f86b6b03e6d87d"
 dependencies = [
- "bigdecimal",
- "either",
- "hashbrown 0.16.1",
- "indexmap",
  "indoc",
- "inventory",
  "libc",
- "lock_api",
  "memoffset",
- "num-bigint",
- "num-complex",
- "num-rational",
- "num-traits",
  "once_cell",
- "ordered-float",
- "parking_lot",
  "portable-atomic",
  "pyo3-build-config",
  "pyo3-ffi",
  "pyo3-macros",
- "rust_decimal",
- "smallvec",
  "unindent",
 ]
 
@@ -3740,12 +3563,6 @@ 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"
@@ -3832,16 +3649,6 @@ dependencies = [
  "tokio",
 ]
 
-[[package]]
-name = "rust_decimal"
-version = "1.39.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "35affe401787a9bd846712274d97654355d21b2a2c092a3139aabe31e9022282"
-dependencies = [
- "arrayvec",
- "num-traits",
-]
-
 [[package]]
 name = "rustc-hash"
 version = "1.1.0"
@@ -4706,24 +4513,12 @@ version = "1.0.22"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9312f7c4f6ff9069b165498234ce8be658059c6728633667c526e27dc2cf1df5"
 
-[[package]]
-name = "unicode-segmentation"
-version = "1.12.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f6ccf251212114b54433ec949fd6a7841275f9ada20dddd2f29e9ceea4501493"
-
 [[package]]
 name = "unicode-width"
 version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b4ac048d71ede7ee76d585517add45da530660ef4390e49b098733c6e897f254"
 
-[[package]]
-name = "unicode-xid"
-version = "0.2.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ebc1c04c71510c7f702b52b7c350734c9ff1295c464a03335b00bb84fc54f853"
-
 [[package]]
 name = "unicode_names2"
 version = "1.3.0"
@@ -4807,15 +4602,6 @@ 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"

Cargo.toml

@@ -26,49 +26,20 @@ opt-level = 3
 networking = { path = "rust/networking" }
 util = { path = "rust/util" }
 
-# Proc-macro authoring tools
-syn = "2.0"
-quote = "1.0"
-proc-macro2 = "1.0"
-darling = "0.20"
-
 # Macro dependecies
 extend = "1.2"
 delegate = "0.13"
-impl-trait-for-tuples = "0.2"
-clap = "4.5"
-derive_more = { version = "2.0.1", features = ["display"] }
-pin-project = "1"
 
 # Utility dependencies
-itertools = "0.14"
-thiserror = "2"
-internment = "0.8"
-recursion = "0.5"
-regex = "1.11"
-once_cell = "1.21"
-thread_local = "1.1"
-bon = "3.4"
-generativity = "1.1"
-anyhow = "1.0"
 keccak-const = "0.2"
 
-# Functional generics/lenses frameworks
-frunk_core = "0.4"
-frunk = "0.4"
-frunk_utils = "0.2"
-frunk-enum-core = "0.3"
-
 # Async dependencies
 tokio = "1.46"
-futures = "0.3"
-futures-util = "0.3"
+futures-lite = "2.6.1"
 futures-timer = "3.0"
 
 # Data structures
 either = "1.15"
-ordered-float = "5.0"
-ahash = "0.8"
 
 # Tracing/logging
 log = "0.4"

MISSED_THINGS.md

@@ -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).

README.md

@@ -72,16 +72,30 @@ There are two ways to run exo:
 
 ### Run from Source (macOS)
 
+If you have [Nix](https://nixos.org/) installed, you can skip most of the steps below and run exo directly:
+
+```bash
+nix run .#exo
+```
+
+**Note:** To accept the Cachix binary cache (and avoid the Xcode Metal ToolChain), add to `/etc/nix/nix.conf`:
+```
+trusted-users = root    (or your username)
+experimental-features = nix-command flakes
+```
+Then restart the Nix daemon: `sudo launchctl kickstart -k system/org.nixos.nix-daemon`
+
 **Prerequisites:**
+- [Xcode](https://developer.apple.com/xcode/) (provides the Metal ToolChain required for MLX compilation)
 - [brew](https://github.com/Homebrew/brew) (for simple package management on macOS)
-  
+
   ```bash
   /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
   ```
 - [uv](https://github.com/astral-sh/uv) (for Python dependency management)
 - [macmon](https://github.com/vladkens/macmon) (for hardware monitoring on Apple Silicon)
 - [node](https://github.com/nodejs/node) (for building the dashboard)
-  
+
   ```bash
   brew install uv macmon node
   ```

TODO.md

@@ -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:

app/EXO/EXO/ContentView.swift

@@ -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 {

app/EXO/EXO/ExoProcessController.swift

@@ -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
@@ -118,11 +126,37 @@ final class ExoProcessController: ObservableObject {
             return
         }
         process.terminationHandler = nil
-        if process.isRunning {
-            process.terminate()
-        }
-        self.process = nil
         status = .stopped
+
+        guard process.isRunning else {
+            self.process = nil
+            return
+        }
+
+        let proc = process
+        self.process = nil
+
+        Task.detached {
+            proc.interrupt()
+
+            for _ in 0..<50 {
+                if !proc.isRunning { return }
+                try? await Task.sleep(nanoseconds: 100_000_000)
+            }
+
+            if proc.isRunning {
+                proc.terminate()
+            }
+
+            for _ in 0..<30 {
+                if !proc.isRunning { return }
+                try? await Task.sleep(nanoseconds: 100_000_000)
+            }
+
+            if proc.isRunning {
+                kill(proc.processIdentifier, SIGKILL)
+            }
+        }
     }
 
     func restart() {

app/EXO/EXO/Models/ClusterState.swift

@@ -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?

app/EXO/EXO/Services/BugReportService.swift

@@ -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 {

app/EXO/EXO/Services/NetworkStatusService.swift

@@ -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 [] }

bench/bench.toml

@@ -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",
+]

bench/exo_bench.py

@@ -1,21 +1,45 @@
+# type: ignore
 #!/usr/bin/env python3
-# pyright: reportAny=false, reportUnknownMemberType=false, reportUnknownVariableType=false, reportUnknownArgumentType=false
+"""Tool-calling eval for exo's OpenAI-compatible API.
+
+Tests whether models correctly:
+- Trigger tool calls when appropriate
+- Return valid JSON arguments matching function schemas
+- Handle multi-turn tool use (call -> result -> final answer)
+- Avoid calling tools when unnecessary
+
+Start exo with a model first, then run:
+    uv run python tool_call_eval.py --model <model-id>
+    uv run python tool_call_eval.py --model <model-id> --host 10.0.0.5 --port 52415
+    uv run python tool_call_eval.py --model <model-id> --repeat 3
+    uv run python tool_call_eval.py --model <model-id> --scenarios weather_simple calculator_multi_turn
+"""
+
 from __future__ import annotations
 
 import argparse
 import contextlib
-import http.client
 import itertools
 import json
-import os
 import sys
 import time
 from collections.abc import Callable
 from pathlib import Path
 from statistics import mean
 from typing import Any
-from urllib.parse import urlencode
 
+from harness import (
+    ExoClient,
+    ExoHttpError,
+    add_common_instance_args,
+    instance_id_from_instance,
+    nodes_used_in_instance,
+    resolve_model_short_id,
+    run_planning_phase,
+    settle_and_fetch_placements,
+    wait_for_instance_gone,
+    wait_for_instance_ready,
+)
 from loguru import logger
 from transformers import AutoTokenizer
 
@@ -98,154 +122,6 @@ def load_tokenizer_for_bench(model_id: str) -> Any:
     return AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
 
 
-class ExoHttpError(RuntimeError):
-    def __init__(self, status: int, reason: str, body_preview: str):
-        super().__init__(f"HTTP {status} {reason}: {body_preview}")
-        self.status = status
-
-
-class ExoClient:
-    def __init__(self, host: str, port: int, timeout_s: float = 7200.0):
-        self.host = host
-        self.port = port
-        self.timeout_s = timeout_s
-
-    def request_json(
-        self,
-        method: str,
-        path: str,
-        params: dict[str, Any] | None = None,
-        body: dict[str, Any] | None = None,
-        headers: dict[str, str] | None = None,
-    ) -> Any:
-        if not path.startswith("/"):
-            path = "/" + path
-        if params:
-            path = path + "?" + urlencode(params)
-
-        conn = http.client.HTTPConnection(self.host, self.port, timeout=self.timeout_s)
-        try:
-            payload: bytes | None = None
-            hdrs: dict[str, str] = {"Accept": "application/json"}
-
-            if body is not None:
-                payload = json.dumps(body).encode("utf-8")
-                hdrs["Content-Type"] = "application/json"
-            if headers:
-                hdrs.update(headers)
-
-            conn.request(method.upper(), path, body=payload, headers=hdrs)
-            resp = conn.getresponse()
-            raw = resp.read()
-            text = raw.decode("utf-8", errors="replace") if raw else ""
-
-            if resp.status >= 400:
-                raise ExoHttpError(resp.status, resp.reason, text[:300])
-
-            if not text:
-                return None
-            return json.loads(text)
-        finally:
-            conn.close()
-
-    def post_bench_chat_completions(self, payload: dict[str, Any]) -> dict[str, Any]:
-        return self.request_json("POST", "/bench/chat/completions", body=payload)
-
-
-def unwrap_instance(instance: dict[str, Any]) -> dict[str, Any]:
-    if len(instance) != 1:
-        raise KeyError(f"Expected 1 key, got keys={list(instance.keys())}")
-
-    tag = next(iter(instance))
-    inner = instance[tag]
-    if not isinstance(inner, dict):
-        raise TypeError(f"payload for {tag} must be dict, got {type(inner)}")
-    return inner
-
-
-def instance_id_from_instance(instance: dict[str, Any]) -> str:
-    inner = unwrap_instance(instance)
-    return str(inner["instanceId"])
-
-
-def nodes_used_in_instance(instance: dict[str, Any]) -> int:
-    inner = unwrap_instance(instance)
-    return len(inner["shardAssignments"]["nodeToRunner"])
-
-
-def runner_ids_from_instance(instance: dict[str, Any]) -> list[str]:
-    inner = unwrap_instance(instance)
-    runner_to_shard = inner["shardAssignments"]["runnerToShard"]
-    return list(runner_to_shard.keys())
-
-
-def runner_ready(runner: dict[str, Any]) -> bool:
-    return "RunnerReady" in runner
-
-
-def runner_failed(runner: dict[str, Any]) -> bool:
-    return "RunnerFailed" in runner
-
-
-def get_runner_failed_message(runner: dict[str, Any]) -> str | None:
-    if "RunnerFailed" in runner:
-        return runner["RunnerFailed"].get("errorMessage")
-    return None
-
-
-def wait_for_instance_ready(
-    client: ExoClient, instance_id: str, timeout: float = 24000.0
-) -> None:
-    start_time = time.time()
-    instance_existed = False
-    while time.time() - start_time < timeout:
-        state = client.request_json("GET", "/state")
-        instances = state.get("instances", {})
-
-        if instance_id not in instances:
-            if instance_existed:
-                # Instance was deleted after being created - likely due to runner failure
-                raise RuntimeError(
-                    f"Instance {instance_id} was deleted (runner may have failed)"
-                )
-            time.sleep(0.1)
-            continue
-
-        instance_existed = True
-        instance = instances[instance_id]
-        runner_ids = runner_ids_from_instance(instance)
-        runners = state.get("runners", {})
-
-        # Check for failed runners first
-        for rid in runner_ids:
-            runner = runners.get(rid, {})
-            if runner_failed(runner):
-                error_msg = get_runner_failed_message(runner) or "Unknown error"
-                raise RuntimeError(f"Runner {rid} failed: {error_msg}")
-
-        if all(runner_ready(runners.get(rid, {})) for rid in runner_ids):
-            return
-
-        time.sleep(0.1)
-
-    raise TimeoutError(f"Instance {instance_id} did not become ready within {timeout=}")
-
-
-def wait_for_instance_gone(
-    client: ExoClient, instance_id: str, timeout: float = 3.0
-) -> None:
-    start_time = time.time()
-    while time.time() - start_time < timeout:
-        try:
-            client.request_json("GET", f"/instance/{instance_id}")
-            time.sleep(0.4)
-        except ExoHttpError as e:
-            if e.status == 404:
-                return
-
-    raise TimeoutError(f"Instance {instance_id} did not get deleted within {timeout=}")
-
-
 def format_peak_memory(b: float) -> str:
     for unit in ["B", "KB", "MB", "GB", "TB"]:
         if b < 1024.0:
@@ -264,39 +140,6 @@ def parse_int_list(values: list[str]) -> list[int]:
     return items
 
 
-def resolve_model_short_id(client: ExoClient, model_arg: str) -> tuple[str, str]:
-    models = client.request_json("GET", "/models") or {}
-    data = models.get("data") or []
-
-    for m in data:
-        if m.get("name").lower() == model_arg.lower():
-            short_id = str(m["name"])
-            full_id = str(m.get("hugging_face_id") or m["name"])
-            return short_id, full_id
-
-    for m in data:
-        if m.get("hugging_face_id") == model_arg:
-            short_id = str(m["name"])
-            full_id = str(m["hugging_face_id"])
-            return short_id, full_id
-
-    raise ValueError(f"Model not found in /models: {model_arg}")
-
-
-def placement_filter(instance_meta: str, wanted: str) -> bool:
-    s = (instance_meta or "").lower()
-    if wanted == "both":
-        return ("ring" in s) or ("jaccl" in s)
-    return wanted in s
-
-
-def sharding_filter(sharding: str, wanted: str) -> bool:
-    s = (sharding or "").lower()
-    if wanted == "both":
-        return ("pipeline" in s) or ("tensor" in s)
-    return wanted in s
-
-
 def run_one_completion(
     client: ExoClient, model_id: str, pp_hint: int, tg: int, prompt_sizer: PromptSizer
 ) -> tuple[dict[str, Any], int]:
@@ -393,11 +236,7 @@ def main() -> int:
         prog="exo-bench",
         description="Benchmark exo model throughput across placement previews.",
     )
-    ap.add_argument("--host", default=os.environ.get("EXO_HOST", "localhost"))
-    ap.add_argument(
-        "--port", type=int, default=int(os.environ.get("EXO_PORT", "52415"))
-    )
-    ap.add_argument("--model", required=True, help="Model short id or huggingface id")
+    add_common_instance_args(ap)
     ap.add_argument(
         "--pp",
         nargs="+",
@@ -410,34 +249,6 @@ def main() -> int:
         required=True,
         help="Generation lengths (ints). Accepts commas.",
     )
-    ap.add_argument(
-        "--max-nodes",
-        type=int,
-        default=4,
-        help="Only consider placements using <= this many nodes.",
-    )
-    ap.add_argument(
-        "--min-nodes",
-        type=int,
-        default=1,
-        help="Only consider placements using >= this many nodes.",
-    )
-    ap.add_argument(
-        "--instance-meta", choices=["ring", "jaccl", "both"], default="both"
-    )
-    ap.add_argument(
-        "--sharding", choices=["pipeline", "tensor", "both"], default="both"
-    )
-    ap.add_argument(
-        "--skip-pipeline-jaccl",
-        action="store_true",
-        help="Skip pipeline+jaccl placements, as it's often pointless.",
-    )
-    ap.add_argument(
-        "--skip-tensor-ring",
-        action="store_true",
-        help="Skip tensor+ring placements, as it's so slow.",
-    )
     ap.add_argument(
         "--repeat", type=int, default=1, help="Repetitions per (pp,tg) pair."
     )
@@ -447,9 +258,6 @@ def main() -> int:
         default=0,
         help="Warmup runs per placement (uses first pp/tg).",
     )
-    ap.add_argument(
-        "--timeout", type=float, default=7200.0, help="HTTP timeout (seconds)."
-    )
     ap.add_argument(
         "--json-out",
         default="bench/results.json",
@@ -487,11 +295,6 @@ 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 +306,9 @@ 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
-
-        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)
+    selected = settle_and_fetch_placements(
+        client, full_model_id, args, settle_timeout=args.settle_timeout
+    )
 
     if not selected:
         logger.error("No valid placements matched your filters.")
@@ -575,6 +333,24 @@ def main() -> int:
     if args.dry_run:
         return 0
 
+    settle_deadline = (
+        time.monotonic() + args.settle_timeout if args.settle_timeout > 0 else None
+    )
+
+    logger.info("Planning phase: checking downloads...")
+    download_duration_s = run_planning_phase(
+        client,
+        full_model_id,
+        selected[0],
+        args.danger_delete_downloads,
+        args.timeout,
+        settle_deadline,
+    )
+    if download_duration_s is not None:
+        logger.info(f"Download: {download_duration_s:.1f}s (freshly downloaded)")
+    else:
+        logger.info("Download: model already cached")
+
     all_rows: list[dict[str, Any]] = []
 
     for preview in selected:
@@ -637,6 +413,11 @@ def main() -> int:
                             "pp_tokens": actual_pp_tokens,
                             "tg": tg,
                             "repeat_index": r,
+                            **(
+                                {"download_duration_s": download_duration_s}
+                                if download_duration_s is not None
+                                else {}
+                            ),
                         }
                     )
                     runs.append(row)

bench/harness.py

@@ -0,0 +1,489 @@
+# type: ignore
+from __future__ import annotations
+
+import argparse
+import http.client
+import json
+import os
+import time
+from typing import Any
+from urllib.parse import urlencode
+
+from loguru import logger
+
+_SETTLE_INITIAL_BACKOFF_S = 1.0
+_SETTLE_MAX_BACKOFF_S = 60.0
+_SETTLE_BACKOFF_MULTIPLIER = 2.0
+
+
+class ExoHttpError(RuntimeError):
+    def __init__(self, status: int, reason: str, body_preview: str):
+        super().__init__(f"HTTP {status} {reason}: {body_preview}")
+        self.status = status
+
+
+class ExoClient:
+    def __init__(self, host: str, port: int, timeout_s: float = 7200.0):
+        self.host = host
+        self.port = port
+        self.timeout_s = timeout_s
+
+    def request_json(
+        self,
+        method: str,
+        path: str,
+        params: dict[str, Any] | None = None,
+        body: dict[str, Any] | None = None,
+        headers: dict[str, str] | None = None,
+    ) -> Any:
+        if not path.startswith("/"):
+            path = "/" + path
+        if params:
+            path = path + "?" + urlencode(params)
+
+        conn = http.client.HTTPConnection(self.host, self.port, timeout=self.timeout_s)
+        try:
+            payload: bytes | None = None
+            hdrs: dict[str, str] = {"Accept": "application/json"}
+
+            if body is not None:
+                payload = json.dumps(body).encode("utf-8")
+                hdrs["Content-Type"] = "application/json"
+            if headers:
+                hdrs.update(headers)
+
+            conn.request(method.upper(), path, body=payload, headers=hdrs)
+            resp = conn.getresponse()
+            raw = resp.read()
+            text = raw.decode("utf-8", errors="replace") if raw else ""
+
+            if resp.status >= 400:
+                raise ExoHttpError(resp.status, resp.reason, text[:300])
+
+            if not text:
+                return None
+            return json.loads(text)
+        finally:
+            conn.close()
+
+    def post_bench_chat_completions(self, payload: dict[str, Any]) -> dict[str, Any]:
+        return self.request_json("POST", "/bench/chat/completions", body=payload)
+
+
+def unwrap_instance(instance: dict[str, Any]) -> dict[str, Any]:
+    if len(instance) != 1:
+        raise KeyError(f"Expected 1 key, got keys={list(instance.keys())}")
+
+    tag = next(iter(instance))
+    inner = instance[tag]
+    if not isinstance(inner, dict):
+        raise TypeError(f"payload for {tag} must be dict, got {type(inner)}")
+    return inner
+
+
+def instance_id_from_instance(instance: dict[str, Any]) -> str:
+    inner = unwrap_instance(instance)
+    return str(inner["instanceId"])
+
+
+def nodes_used_in_instance(instance: dict[str, Any]) -> int:
+    inner = unwrap_instance(instance)
+    return len(inner["shardAssignments"]["nodeToRunner"])
+
+
+def runner_ids_from_instance(instance: dict[str, Any]) -> list[str]:
+    inner = unwrap_instance(instance)
+    runner_to_shard = inner["shardAssignments"]["runnerToShard"]
+    return list(runner_to_shard.keys())
+
+
+def runner_ready(runner: dict[str, Any]) -> bool:
+    return "RunnerReady" in runner
+
+
+def runner_failed(runner: dict[str, Any]) -> bool:
+    return "RunnerFailed" in runner
+
+
+def get_runner_failed_message(runner: dict[str, Any]) -> str | None:
+    if "RunnerFailed" in runner:
+        return runner["RunnerFailed"].get("errorMessage")
+    return None
+
+
+def wait_for_instance_ready(
+    client: ExoClient, instance_id: str, timeout: float = 24000.0
+) -> None:
+    start_time = time.time()
+    instance_existed = False
+    while time.time() - start_time < timeout:
+        state = client.request_json("GET", "/state")
+        instances = state.get("instances", {})
+
+        if instance_id not in instances:
+            if instance_existed:
+                # Instance was deleted after being created - likely due to runner failure
+                raise RuntimeError(
+                    f"Instance {instance_id} was deleted (runner may have failed)"
+                )
+            time.sleep(0.1)
+            continue
+
+        instance_existed = True
+        instance = instances[instance_id]
+        runner_ids = runner_ids_from_instance(instance)
+        runners = state.get("runners", {})
+
+        # Check for failed runners first
+        for rid in runner_ids:
+            runner = runners.get(rid, {})
+            if runner_failed(runner):
+                error_msg = get_runner_failed_message(runner) or "Unknown error"
+                raise RuntimeError(f"Runner {rid} failed: {error_msg}")
+
+        if all(runner_ready(runners.get(rid, {})) for rid in runner_ids):
+            return
+
+        time.sleep(0.1)
+
+    raise TimeoutError(f"Instance {instance_id} did not become ready within {timeout=}")
+
+
+def wait_for_instance_gone(
+    client: ExoClient, instance_id: str, timeout: float = 3.0
+) -> None:
+    start_time = time.time()
+    while time.time() - start_time < timeout:
+        try:
+            client.request_json("GET", f"/instance/{instance_id}")
+            time.sleep(0.4)
+        except ExoHttpError as e:
+            if e.status == 404:
+                return
+            raise
+
+    raise TimeoutError(f"Instance {instance_id} did not get deleted within {timeout=}")
+
+
+def resolve_model_short_id(client: ExoClient, model_arg: str) -> tuple[str, str]:
+    models = client.request_json("GET", "/models") or {}
+    data = models.get("data") or []
+
+    for m in data:
+        if (m.get("name") or "").lower() == model_arg.lower():
+            short_id = str(m["name"])
+            full_id = str(m.get("hugging_face_id") or m["name"])
+            return short_id, full_id
+
+    for m in data:
+        if m.get("hugging_face_id") == model_arg:
+            short_id = str(m["name"])
+            full_id = str(m["hugging_face_id"])
+            return short_id, full_id
+
+    raise ValueError(f"Model not found in /models: {model_arg}")
+
+
+def placement_filter(instance_meta: str, wanted: str) -> bool:
+    s = (instance_meta or "").lower()
+    if wanted == "both":
+        return ("ring" in s) or ("jaccl" in s)
+    return wanted in s
+
+
+def sharding_filter(sharding: str, wanted: str) -> bool:
+    s = (sharding or "").lower()
+    if wanted == "both":
+        return ("pipeline" in s) or ("tensor" in s)
+    return wanted in s
+
+
+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 settle_and_fetch_placements(
+    client: ExoClient,
+    full_model_id: str,
+    args: argparse.Namespace,
+    settle_timeout: float = 0,
+) -> list[dict[str, Any]]:
+    selected = fetch_and_filter_placements(client, full_model_id, args)
+
+    if not selected and settle_timeout > 0:
+        backoff = _SETTLE_INITIAL_BACKOFF_S
+        deadline = time.monotonic() + settle_timeout
+        while not selected and time.monotonic() < deadline:
+            remaining = 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)..."
+            )
+            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)
+
+    return selected
+
+
+def run_planning_phase(
+    client: ExoClient,
+    full_model_id: str,
+    preview: dict[str, Any],
+    danger_delete: bool,
+    timeout: float,
+    settle_deadline: float | None,
+) -> float | None:
+    """Check disk space and ensure model is downloaded before benchmarking.
+
+    Returns the wall-clock download duration in seconds if a fresh download
+    was needed, or None if the model was already cached on all nodes.
+    """
+    # 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 None
+
+    # 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", {})
+
+    needs_download = False
+
+    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
+
+        needs_download = True
+
+        # 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 (skip read-only models from EXO_MODELS_PATH)
+        completed = [
+            (
+                unwrap_instance(p["DownloadCompleted"]["shardMetadata"])["modelCard"][
+                    "modelId"
+                ],
+                p["DownloadCompleted"]["totalBytes"]["inBytes"],
+            )
+            for p in node_downloads
+            if "DownloadCompleted" in p
+            and not p["DownloadCompleted"].get("readOnly", False)
+        ]
+        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)
+    download_t0 = time.perf_counter() if needs_download else None
+    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:
+            if download_t0 is not None:
+                return time.perf_counter() - download_t0
+            return None
+        time.sleep(1)
+
+    raise TimeoutError("Downloads did not complete in time")
+
+
+def add_common_instance_args(ap: argparse.ArgumentParser) -> None:
+    ap.add_argument("--host", default=os.environ.get("EXO_HOST", "localhost"))
+    ap.add_argument(
+        "--port", type=int, default=int(os.environ.get("EXO_PORT", "52415"))
+    )
+    ap.add_argument("--model", required=True, help="Model short id or huggingface id")
+    ap.add_argument(
+        "--max-nodes",
+        type=int,
+        default=4,
+        help="Only consider placements using <= this many nodes.",
+    )
+    ap.add_argument(
+        "--min-nodes",
+        type=int,
+        default=1,
+        help="Only consider placements using >= this many nodes.",
+    )
+    ap.add_argument(
+        "--instance-meta", choices=["ring", "jaccl", "both"], default="both"
+    )
+    ap.add_argument(
+        "--sharding", choices=["pipeline", "tensor", "both"], default="both"
+    )
+    ap.add_argument(
+        "--skip-pipeline-jaccl",
+        action="store_true",
+        help="Skip pipeline+jaccl placements, as it's often pointless.",
+    )
+    ap.add_argument(
+        "--skip-tensor-ring",
+        action="store_true",
+        help="Skip tensor+ring placements, as it's so slow.",
+    )
+    ap.add_argument(
+        "--timeout", type=float, default=7200.0, help="HTTP timeout (seconds)."
+    )
+    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.",
+    )

bench/pyproject.toml

@@ -0,0 +1,17 @@
+[project]
+name = "exo-bench"
+version = "0.1.0"
+description = "Benchmarking tool for exo distributed inference"
+requires-python = ">=3.13"
+dependencies = [
+    "httpx>=0.27.0",
+    "loguru>=0.7.3",
+    "transformers>=5.0.0",
+    "huggingface-hub>=0.33.4",
+    "tiktoken>=0.12.0",
+    "jinja2>=3.1.0",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"

bench/scenarios.toml

@@ -0,0 +1,306 @@
+# Tool definitions — each becomes an OpenAI function tool.
+# All scenarios get all tools unless they specify a `tools` list.
+
+[tools.get_current_weather]
+description = "Get the current weather in a given location"
+required = ["location"]
+
+[tools.get_current_weather.properties.location]
+type = "string"
+description = "City and state, e.g. San Francisco, CA"
+
+[tools.get_current_weather.properties.unit]
+type = "string"
+enum = ["celsius", "fahrenheit"]
+description = "Temperature unit"
+
+[tools.calculate]
+description = "Evaluate a mathematical expression and return the numeric result"
+required = ["expression"]
+
+[tools.calculate.properties.expression]
+type = "string"
+description = "The math expression to evaluate, e.g. '2 + 3 * 4'"
+
+[tools.search_products]
+description = "Search for products in a catalog by query, category, and price"
+required = ["query"]
+
+[tools.search_products.properties.query]
+type = "string"
+description = "Search query string"
+
+[tools.search_products.properties.category]
+type = "string"
+enum = ["electronics", "clothing", "food", "books"]
+description = "Product category to filter by"
+
+[tools.search_products.properties.max_price]
+type = "number"
+description = "Maximum price in USD"
+
+[tools.create_todos]
+description = "Create a structured todo list"
+required = ["todos"]
+
+[tools.create_todos.properties.todos]
+type = "array"
+description = "List of todo items"
+
+[tools.create_todos.properties.todos.items]
+type = "object"
+required = ["content", "status", "priority"]
+
+[tools.create_todos.properties.todos.items.properties.content]
+type = "string"
+description = "The todo item text"
+
+[tools.create_todos.properties.todos.items.properties.status]
+type = "string"
+description = "Status: pending, in_progress, or completed"
+
+[tools.create_todos.properties.todos.items.properties.priority]
+type = "string"
+description = "Priority: low, normal, or high"
+
+# -- Should call a tool --
+
+[[scenarios]]
+name = "weather_simple"
+description = "Basic weather query -> get_current_weather"
+expect_tool_call = true
+expected_function = "get_current_weather"
+required_arg_keys = ["location"]
+
+[[scenarios.messages]]
+role = "user"
+content = "What's the weather like in Tokyo right now?"
+
+[[scenarios]]
+name = "calculator_simple"
+description = "Math question -> calculate"
+expect_tool_call = true
+expected_function = "calculate"
+required_arg_keys = ["expression"]
+
+[[scenarios.messages]]
+role = "user"
+content = "Use the calculator to compute 3847 * 926 + 17293"
+
+[[scenarios]]
+name = "search_with_filters"
+description = "Product search with category and price filter"
+expect_tool_call = true
+expected_function = "search_products"
+required_arg_keys = ["query"]
+
+[[scenarios.messages]]
+role = "user"
+content = "Find me electronics under $50"
+
+# -- Multi-turn: tool call then follow-up --
+
+[[scenarios]]
+name = "weather_multi_turn"
+description = "Weather query -> tool result -> natural language summary"
+expect_tool_call = true
+expected_function = "get_current_weather"
+required_arg_keys = ["location"]
+
+[scenarios.tool_result]
+temperature = "18C"
+condition = "partly cloudy"
+humidity = "65%"
+wind = "12 km/h NW"
+
+[[scenarios.messages]]
+role = "user"
+content = "What's the weather in Paris?"
+
+[[scenarios]]
+name = "calculator_multi_turn"
+description = "Math query -> tool result -> model reports the answer"
+expect_tool_call = true
+expected_function = "calculate"
+required_arg_keys = ["expression"]
+
+[scenarios.tool_result]
+result = 491682
+
+[[scenarios.messages]]
+role = "user"
+content = "Use the calculator to compute 1847 * 263 + 5921"
+
+[[scenarios]]
+name = "search_multi_turn"
+description = "Search query -> tool result -> model summarizes products"
+expect_tool_call = true
+expected_function = "search_products"
+required_arg_keys = ["query"]
+
+[[scenarios.tool_result.results]]
+name = "Hands-On Machine Learning"
+price = 45.99
+rating = 4.8
+
+[[scenarios.tool_result.results]]
+name = "Deep Learning with Python"
+price = 39.99
+rating = 4.6
+
+[[scenarios.messages]]
+role = "user"
+content = "Search for books about machine learning"
+
+# -- Sequential tool calls --
+
+[[scenarios]]
+name = "chained_tool_calls_same"
+description = "Thinking + weather(Tokyo) -> result -> model must call weather(London)"
+expect_tool_call = true
+expected_function = "get_current_weather"
+required_arg_keys = ["location"]
+
+[[scenarios.messages]]
+role = "user"
+content = "Compare the weather in Tokyo and London."
+
+[[scenarios.messages]]
+role = "assistant"
+content = "I'll check both cities. Let me start with Tokyo."
+
+[[scenarios.messages.tool_calls]]
+id = "call_1"
+name = "get_current_weather"
+arguments = { location = "Tokyo" }
+
+[[scenarios.messages]]
+role = "tool"
+tool_call_id = "call_1"
+content = '{"temperature": "25C", "condition": "sunny"}'
+
+[[scenarios]]
+name = "chained_tool_calls_different"
+description = "Thinking + weather(Berlin) -> result -> model must call calculator"
+expect_tool_call = true
+expected_function = "calculate"
+required_arg_keys = ["expression"]
+
+[[scenarios.messages]]
+role = "user"
+content = "What's the weather in Berlin, and also use the calculator to compute 4819 * 37 + 291."
+
+[[scenarios.messages]]
+role = "assistant"
+content = "I'll handle both. Let me check Berlin's weather first."
+
+[[scenarios.messages.tool_calls]]
+id = "call_2"
+name = "get_current_weather"
+arguments = { location = "Berlin" }
+
+[[scenarios.messages]]
+role = "tool"
+tool_call_id = "call_2"
+content = '{"temperature": "12C", "condition": "rainy"}'
+
+[[scenarios]]
+name = "chained_tool_calls_three"
+description = "Two prior thinking+tool calls -> results -> model must make a third"
+expect_tool_call = true
+expected_function = "get_current_weather"
+required_arg_keys = ["location"]
+
+[[scenarios.messages]]
+role = "user"
+content = "Compare weather in Tokyo, Paris, and London."
+
+[[scenarios.messages]]
+role = "assistant"
+content = "I'll check all three cities. Starting with Tokyo."
+
+[[scenarios.messages.tool_calls]]
+id = "call_3"
+name = "get_current_weather"
+arguments = { location = "Tokyo" }
+
+[[scenarios.messages]]
+role = "tool"
+tool_call_id = "call_3"
+content = '{"temperature": "25C", "condition": "sunny"}'
+
+[[scenarios.messages]]
+role = "assistant"
+content = "Got Tokyo. Now checking Paris."
+
+[[scenarios.messages.tool_calls]]
+id = "call_4"
+name = "get_current_weather"
+arguments = { location = "Paris" }
+
+[[scenarios.messages]]
+role = "tool"
+tool_call_id = "call_4"
+content = '{"temperature": "18C", "condition": "cloudy"}'
+
+# -- Nested object schema (regression for lossy chat template rendering) --
+
+[[scenarios]]
+name = "nested_schema_tool_call"
+description = "Tool call with nested object array schema -> create_todos"
+expect_tool_call = true
+expected_function = "create_todos"
+required_arg_keys = ["todos"]
+nested_array_key = "todos"
+required_item_keys = ["content", "status", "priority"]
+tools = ["create_todos"]
+
+[[scenarios.messages]]
+role = "user"
+content = "Create a todo list with 3 items to learn Python"
+
+# -- Tool name integrity (regression for harmony token leaking into name) --
+
+[tools.glob]
+description = "Search for files matching a glob pattern in the codebase"
+required = ["pattern"]
+
+[tools.glob.properties.pattern]
+type = "string"
+description = "The glob pattern to match files against, e.g. '**/*.py'"
+
+[tools.glob.properties.path]
+type = "string"
+description = "The directory to search in"
+
+[[scenarios]]
+name = "tool_name_integrity"
+description = "Tool name must not contain harmony tokens like <|channel|>"
+expect_tool_call = true
+expected_function = "glob"
+required_arg_keys = ["pattern"]
+tools = ["glob"]
+
+[[scenarios.messages]]
+role = "user"
+content = "Find all Python files in the src directory"
+
+# -- Should NOT call a tool --
+
+[[scenarios]]
+name = "no_tool_joke"
+description = "Joke request should NOT trigger any tool"
+expect_tool_call = false
+
+[[scenarios.messages]]
+role = "user"
+content = "Tell me a funny joke about cats."
+
+[[scenarios]]
+name = "no_tool_factual"
+description = "Factual question answerable from training data"
+expect_tool_call = false
+
+[[scenarios.messages]]
+role = "user"
+content = "What is the capital of Japan?"

bench/single-m3-ultra.toml

@@ -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))"]

dashboard/src/lib/components/ChatForm.svelte

@@ -12,6 +12,9 @@
     ttftMs,
     tps,
     totalTokens,
+    thinkingEnabled as thinkingEnabledStore,
+    setConversationThinking,
+    stopGeneration,
   } from "$lib/stores/app.svelte";
   import ChatAttachments from "./ChatAttachments.svelte";
   import ImageParamsPanel from "./ImageParamsPanel.svelte";
@@ -25,6 +28,7 @@
     autofocus?: boolean;
     showModelSelector?: boolean;
     modelTasks?: Record<string, string[]>;
+    modelCapabilities?: Record<string, string[]>;
   }
 
   let {
@@ -34,6 +38,7 @@
     autofocus = true,
     showModelSelector = false,
     modelTasks = {},
+    modelCapabilities = {},
   }: Props = $props();
 
   let message = $state("");
@@ -41,6 +46,7 @@
   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 +101,12 @@
     );
   });
 
+  const modelSupportsThinking = $derived(() => {
+    if (!currentModel) return false;
+    const caps = modelCapabilities[currentModel] || [];
+    return caps.includes("thinking_toggle") && caps.includes("text");
+  });
+
   const isEditOnlyWithoutImage = $derived(
     currentModel !== null &&
       modelSupportsOnlyImageEditing(currentModel) &&
@@ -254,6 +266,7 @@
 
   function handleSubmit() {
     if ((!message.trim() && uploadedFiles.length === 0) || loading) return;
+    if (isEditOnlyWithoutImage) return;
 
     const content = message.trim();
     const files = [...uploadedFiles];
@@ -278,11 +291,19 @@
       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 +541,35 @@
             </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">
@@ -599,92 +649,97 @@
             : 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>
 
-      <button
-        type="submit"
-        disabled={!canSend || loading || isEditOnlyWithoutImage}
-        class="px-2.5 sm:px-4 py-1.5 sm:py-2 rounded text-xs sm:text-xs tracking-[0.1em] sm:tracking-[0.15em] uppercase font-medium transition-all duration-200 whitespace-nowrap
-					{!canSend || loading || isEditOnlyWithoutImage
-          ? 'bg-exo-medium-gray/50 text-exo-light-gray cursor-not-allowed'
-          : 'bg-exo-yellow text-exo-black hover:bg-exo-yellow-darker hover:shadow-[0_0_20px_rgba(255,215,0,0.3)]'}"
-        aria-label={shouldShowEditMode
-          ? "Edit image"
-          : isImageModel()
-            ? "Generate image"
-            : "Send message"}
-      >
-        {#if loading}
+      {#if loading}
+        <button
+          type="button"
+          onclick={() => stopGeneration()}
+          class="px-2.5 sm:px-4 py-1.5 sm:py-2 rounded text-xs sm:text-xs tracking-[0.1em] sm:tracking-[0.15em] font-medium transition-all duration-200 whitespace-nowrap bg-exo-medium-gray/70 text-exo-light-gray hover:bg-exo-medium-gray hover:text-white"
+          aria-label="Stop generation"
+        >
           <span class="inline-flex items-center gap-1 sm:gap-2">
-            <span
-              class="w-2.5 h-2.5 sm:w-3 sm:h-3 border-2 border-current border-t-transparent rounded-full animate-spin"
-            ></span>
-            <span class="hidden sm:inline"
-              >{shouldShowEditMode
-                ? "EDITING"
-                : isImageModel()
-                  ? "GENERATING"
-                  : "PROCESSING"}</span
-            >
-            <span class="sm:hidden">...</span>
-          </span>
-        {:else if shouldShowEditMode}
-          <span class="inline-flex items-center gap-1.5">
             <svg
-              class="w-3.5 h-3.5"
-              fill="none"
+              class="w-3 h-3 sm:w-3.5 sm:h-3.5"
+              fill="currentColor"
               viewBox="0 0 24 24"
-              stroke="currentColor"
-              stroke-width="2"
             >
-              <path
-                stroke-linecap="round"
-                stroke-linejoin="round"
-                d="M11 5H6a2 2 0 00-2 2v11a2 2 0 002 2h11a2 2 0 002-2v-5m-1.414-9.414a2 2 0 112.828 2.828L11.828 15H9v-2.828l8.586-8.586z"
-              />
+              <rect x="6" y="6" width="12" height="12" rx="1" />
             </svg>
-            <span>EDIT</span>
+            <span class="hidden sm:inline">Cancel</span>
           </span>
-        {:else if isEditOnlyWithoutImage}
-          <span class="inline-flex items-center gap-1.5">
-            <svg
-              class="w-3.5 h-3.5"
-              fill="none"
-              viewBox="0 0 24 24"
-              stroke="currentColor"
-              stroke-width="2"
-            >
-              <path
-                stroke-linecap="round"
-                stroke-linejoin="round"
-                d="M11 5H6a2 2 0 00-2 2v11a2 2 0 002 2h11a2 2 0 002-2v-5m-1.414-9.414a2 2 0 112.828 2.828L11.828 15H9v-2.828l8.586-8.586z"
-              />
-            </svg>
-            <span>EDIT</span>
-          </span>
-        {:else if isImageModel()}
-          <span class="inline-flex items-center gap-1.5">
-            <svg
-              class="w-3.5 h-3.5"
-              fill="none"
-              viewBox="0 0 24 24"
-              stroke="currentColor"
-              stroke-width="2"
-            >
-              <rect x="3" y="3" width="18" height="18" rx="2" ry="2" />
-              <circle cx="8.5" cy="8.5" r="1.5" />
-              <polyline points="21 15 16 10 5 21" />
-            </svg>
-            <span>GENERATE</span>
-          </span>
-        {:else}
-          SEND
-        {/if}
-      </button>
+        </button>
+      {:else}
+        <button
+          type="submit"
+          disabled={!canSend || isEditOnlyWithoutImage}
+          class="px-2.5 sm:px-4 py-1.5 sm:py-2 rounded text-xs sm:text-xs tracking-[0.1em] sm:tracking-[0.15em] uppercase font-medium transition-all duration-200 whitespace-nowrap
+					{!canSend || isEditOnlyWithoutImage
+            ? 'bg-exo-medium-gray/50 text-exo-light-gray cursor-not-allowed'
+            : 'bg-exo-yellow text-exo-black hover:bg-exo-yellow-darker hover:shadow-[0_0_20px_rgba(255,215,0,0.3)]'}"
+          aria-label={shouldShowEditMode
+            ? "Edit image"
+            : isImageModel()
+              ? "Generate image"
+              : "Send message"}
+        >
+          {#if shouldShowEditMode}
+            <span class="inline-flex items-center gap-1.5">
+              <svg
+                class="w-3.5 h-3.5"
+                fill="none"
+                viewBox="0 0 24 24"
+                stroke="currentColor"
+                stroke-width="2"
+              >
+                <path
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                  d="M11 5H6a2 2 0 00-2 2v11a2 2 0 002 2h11a2 2 0 002-2v-5m-1.414-9.414a2 2 0 112.828 2.828L11.828 15H9v-2.828l8.586-8.586z"
+                />
+              </svg>
+              <span>EDIT</span>
+            </span>
+          {:else if isEditOnlyWithoutImage}
+            <span class="inline-flex items-center gap-1.5">
+              <svg
+                class="w-3.5 h-3.5"
+                fill="none"
+                viewBox="0 0 24 24"
+                stroke="currentColor"
+                stroke-width="2"
+              >
+                <path
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                  d="M11 5H6a2 2 0 00-2 2v11a2 2 0 002 2h11a2 2 0 002-2v-5m-1.414-9.414a2 2 0 112.828 2.828L11.828 15H9v-2.828l8.586-8.586z"
+                />
+              </svg>
+              <span>EDIT</span>
+            </span>
+          {:else if isImageModel()}
+            <span class="inline-flex items-center gap-1.5">
+              <svg
+                class="w-3.5 h-3.5"
+                fill="none"
+                viewBox="0 0 24 24"
+                stroke="currentColor"
+                stroke-width="2"
+              >
+                <rect x="3" y="3" width="18" height="18" rx="2" ry="2" />
+                <circle cx="8.5" cy="8.5" r="1.5" />
+                <polyline points="21 15 16 10 5 21" />
+              </svg>
+              <span>GENERATE</span>
+            </span>
+          {:else}
+            SEND
+          {/if}
+        </button>
+      {/if}
     </div>
 
     <!-- Bottom accent line -->

dashboard/src/lib/components/ChatMessages.svelte

@@ -3,16 +3,18 @@
     messages,
     currentResponse,
     isLoading,
+    prefillProgress,
     deleteMessage,
     editAndRegenerate,
     regenerateLastResponse,
     regenerateFromToken,
     setEditingImage,
   } from "$lib/stores/app.svelte";
-  import type { Message } from "$lib/stores/app.svelte";
   import type { MessageAttachment } from "$lib/stores/app.svelte";
   import MarkdownContent from "./MarkdownContent.svelte";
   import TokenHeatmap from "./TokenHeatmap.svelte";
+  import PrefillProgressBar from "./PrefillProgressBar.svelte";
+  import ImageLightbox from "./ImageLightbox.svelte";
 
   interface Props {
     class?: string;
@@ -24,6 +26,7 @@
   const messageList = $derived(messages());
   const response = $derived(currentResponse());
   const loading = $derived(isLoading());
+  const prefill = $derived(prefillProgress());
 
   // Scroll management - user controls scroll, show button when not at bottom
   const SCROLL_THRESHOLD = 100;
@@ -101,6 +104,9 @@
   let copiedMessageId = $state<string | null>(null);
   let expandedThinkingMessageIds = $state<Set<string>>(new Set());
 
+  // Lightbox state
+  let expandedImageSrc = $state<string | null>(null);
+
   // Uncertainty heatmap toggle
   let heatmapMessageIds = $state<Set<string>>(new Set());
 
@@ -221,6 +227,7 @@
   }
 
   function handleDeleteClick(messageId: string) {
+    if (loading) return;
     deleteConfirmId = messageId;
   }
 
@@ -251,7 +258,7 @@
 </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'
@@ -313,9 +320,11 @@
           <!-- 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
@@ -389,10 +398,15 @@
                         class="flex items-center gap-2 bg-exo-dark-gray/60 border border-exo-yellow/20 rounded px-2 py-1 text-xs font-mono"
                       >
                         {#if attachment.type === "image" && attachment.preview}
+                          <!-- svelte-ignore a11y_no_noninteractive_element_interactions, a11y_click_events_have_key_events -->
                           <img
                             src={attachment.preview}
                             alt={attachment.name}
-                            class="w-12 h-12 object-cover rounded border border-exo-yellow/20"
+                            class="w-12 h-12 object-cover rounded border border-exo-yellow/20 cursor-pointer hover:border-exo-yellow/50 transition-colors"
+                            onclick={() => {
+                              if (attachment.preview)
+                                expandedImageSrc = attachment.preview;
+                            }}
                           />
                         {:else}
                           <span>{getAttachmentIcon(attachment)}</span>
@@ -416,6 +430,9 @@
             {:else}
               <!-- Assistant message styling -->
               <div class="p-3 sm:p-4">
+                {#if loading && isLastAssistantMessage(message.id) && prefill && !message.content}
+                  <PrefillProgressBar progress={prefill} class="mb-3" />
+                {/if}
                 {#if message.thinking && message.thinking.trim().length > 0}
                   <div
                     class="mb-3 rounded border border-exo-yellow/20 bg-exo-black/40"
@@ -466,15 +483,44 @@
                   <div class="mb-3">
                     {#each message.attachments.filter((a) => a.type === "generated-image") as attachment}
                       <div class="relative group/img inline-block">
+                        <!-- svelte-ignore a11y_no_noninteractive_element_interactions, a11y_click_events_have_key_events -->
                         <img
                           src={attachment.preview}
                           alt=""
-                          class="max-w-full max-h-[512px] rounded-lg border border-exo-yellow/20 shadow-lg shadow-black/20"
+                          class="max-w-full max-h-[512px] rounded-lg border border-exo-yellow/20 shadow-lg shadow-black/20 cursor-pointer"
+                          onclick={() => {
+                            if (attachment.preview)
+                              expandedImageSrc = attachment.preview;
+                          }}
                         />
                         <!-- Button overlay -->
                         <div
                           class="absolute top-2 right-2 flex gap-1 opacity-0 group-hover/img:opacity-100 transition-opacity"
                         >
+                          <!-- Expand button -->
+                          <button
+                            type="button"
+                            class="p-2 rounded-lg bg-exo-dark-gray/80 border border-exo-yellow/30 text-exo-yellow hover:bg-exo-dark-gray hover:border-exo-yellow/50 cursor-pointer"
+                            onclick={() => {
+                              if (attachment.preview)
+                                expandedImageSrc = attachment.preview;
+                            }}
+                            title="Expand image"
+                          >
+                            <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="M4 8V4m0 0h4M4 4l5 5m11-1V4m0 0h-4m4 0l-5 5M4 16v4m0 0h4m-4 0l5-5m11 5l-5-5m5 5v-4m0 4h-4"
+                              />
+                            </svg>
+                          </button>
                           <!-- Edit button -->
                           <button
                             type="button"
@@ -713,8 +759,13 @@
             <!-- 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"
@@ -789,3 +840,8 @@
     </button>
   {/if}
 </div>
+
+<ImageLightbox
+  src={expandedImageSrc}
+  onclose={() => (expandedImageSrc = null)}
+/>

dashboard/src/lib/components/FamilyLogos.svelte

@@ -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
@@ -58,6 +64,18 @@
       d="M17.962 1.844h-4.326l-3.425 7.81H5.369V1.878H1.5V22h3.87v-8.477h6.824a3.025 3.025 0 002.743-1.75V22h3.87v-8.477a3.87 3.87 0 00-3.588-3.86v-.01h-2.125a3.94 3.94 0 002.323-2.12l2.545-5.689z"
     />
   </svg>
+{:else if family === "flux"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M12 2L2 19h7.5l2.5-4.5L14.5 19H22L12 2zm0 4.5L16.5 17h-3l-1.5-2.7L10.5 17h-3L12 6.5z"
+    />
+  </svg>
+{:else if family === "qwen-image"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M12.604 1.34c.393.69.784 1.382 1.174 2.075a.18.18 0 00.157.091h5.552c.174 0 .322.11.446.327l1.454 2.57c.19.337.24.478.024.837-.26.43-.513.864-.76 1.3l-.367.658c-.106.196-.223.28-.04.512l2.652 4.637c.172.301.111.494-.043.77-.437.785-.882 1.564-1.335 2.34-.159.272-.352.375-.68.37-.777-.016-1.552-.01-2.327.016a.099.099 0 00-.081.05 575.097 575.097 0 01-2.705 4.74c-.169.293-.38.363-.725.364-.997.003-2.002.004-3.017.002a.537.537 0 01-.465-.271l-1.335-2.323a.09.09 0 00-.083-.049H4.982c-.285.03-.553-.001-.805-.092l-1.603-2.77a.543.543 0 01-.002-.54l1.207-2.12a.198.198 0 000-.197 550.951 550.951 0 01-1.875-3.272l-.79-1.395c-.16-.31-.173-.496.095-.965.465-.813.927-1.625 1.387-2.436.132-.234.304-.334.584-.335a338.3 338.3 0 012.589-.001.124.124 0 00.107-.063l2.806-4.895a.488.488 0 01.422-.246c.524-.001 1.053 0 1.583-.006L11.704 1c.341-.003.724.032.9.34zm-3.432.403a.06.06 0 00-.052.03L6.254 6.788a.157.157 0 01-.135.078H3.253c-.056 0-.07.025-.041.074l5.81 10.156c.025.042.013.062-.034.063l-2.795.015a.218.218 0 00-.2.116l-1.32 2.31c-.044.078-.021.118.068.118l5.716.008c.046 0 .08.02.104.061l1.403 2.454c.046.081.092.082.139 0l5.006-8.76.783-1.382a.055.055 0 01.096 0l1.424 2.53a.122.122 0 00.107.062l2.763-.02a.04.04 0 00.035-.02.041.041 0 000-.04l-2.9-5.086a.108.108 0 010-.113l.293-.507 1.12-1.977c.024-.041.012-.062-.035-.062H9.2c-.059 0-.073-.026-.043-.077l1.434-2.505a.107.107 0 000-.114L9.225 1.774a.06.06 0 00-.053-.031zm6.29 8.02c.046 0 .058.02.034.06l-.832 1.465-2.613 4.585a.056.056 0 01-.05.029.058.058 0 01-.05-.029L8.498 9.841c-.02-.034-.01-.052.028-.054l.216-.012 6.722-.012z"
+    />
+  </svg>
 {:else if family === "huggingface"}
   <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
     <path

dashboard/src/lib/components/FamilySidebar.svelte

@@ -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",
@@ -22,6 +29,8 @@
     glm: "GLM",
     minimax: "MiniMax",
     kimi: "Kimi",
+    flux: "FLUX",
+    "qwen-image": "Qwen Img",
   };
 
   function getFamilyName(family: string): string {
@@ -32,7 +41,7 @@
 </script>
 
 <div
-  class="flex flex-col gap-1 py-2 px-1 border-r border-exo-yellow/10 bg-exo-medium-gray/30 min-w-[64px]"
+  class="flex flex-col gap-1 py-2 px-1 border-r border-exo-yellow/10 bg-exo-medium-gray/30 min-w-[64px] overflow-y-auto scrollbar-hide"
 >
   <!-- All models (no filter) -->
   <button
@@ -87,6 +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"

dashboard/src/lib/components/HuggingFaceResultItem.svelte

@@ -26,7 +26,8 @@
     downloadedOnNodes = [],
   }: HuggingFaceResultItemProps = $props();
 
-  function formatNumber(num: number): string {
+  function formatNumber(num: number | undefined): string {
+    if (num == null) return "0";
     if (num >= 1000000) {
       return `${(num / 1000000).toFixed(1)}M`;
     } else if (num >= 1000) {

dashboard/src/lib/components/ImageLightbox.svelte

@@ -0,0 +1,96 @@
+<script lang="ts">
+  import { fade, fly } from "svelte/transition";
+  import { cubicOut } from "svelte/easing";
+
+  interface Props {
+    src: string | null;
+    onclose: () => void;
+  }
+
+  let { src, onclose }: Props = $props();
+
+  function handleKeydown(e: KeyboardEvent) {
+    if (e.key === "Escape") {
+      onclose();
+    }
+  }
+
+  function extensionFromSrc(dataSrc: string): string {
+    const match = dataSrc.match(/^data:image\/(\w+)/);
+    if (match) return match[1] === "jpeg" ? "jpg" : match[1];
+    const urlMatch = dataSrc.match(/\.(\w+)(?:\?|$)/);
+    if (urlMatch) return urlMatch[1];
+    return "png";
+  }
+
+  function handleDownload(e: MouseEvent) {
+    e.stopPropagation();
+    if (!src) return;
+    const link = document.createElement("a");
+    link.href = src;
+    link.download = `image-${Date.now()}.${extensionFromSrc(src)}`;
+    link.click();
+  }
+
+  function handleClose(e: MouseEvent) {
+    e.stopPropagation();
+    onclose();
+  }
+</script>
+
+<svelte:window onkeydown={src ? handleKeydown : undefined} />
+
+{#if src}
+  <div
+    class="fixed inset-0 z-50 bg-black/90 backdrop-blur-sm flex items-center justify-center"
+    transition:fade={{ duration: 200 }}
+    onclick={onclose}
+    role="presentation"
+    onintrostart={() => (document.body.style.overflow = "hidden")}
+    onoutroend={() => (document.body.style.overflow = "")}
+  >
+    <div class="absolute top-4 right-4 flex gap-2 z-10">
+      <button
+        type="button"
+        class="p-2 rounded-lg bg-exo-dark-gray/80 border border-exo-yellow/30 text-exo-yellow hover:bg-exo-dark-gray hover:border-exo-yellow/50 cursor-pointer transition-colors"
+        onclick={handleDownload}
+        title="Download image"
+      >
+        <svg
+          class="w-5 h-5"
+          fill="none"
+          viewBox="0 0 24 24"
+          stroke="currentColor"
+          stroke-width="2"
+        >
+          <path
+            stroke-linecap="round"
+            stroke-linejoin="round"
+            d="M4 16v1a3 3 0 003 3h10a3 3 0 003-3v-1m-4-4l-4 4m0 0l-4-4m4 4V4"
+          />
+        </svg>
+      </button>
+      <button
+        type="button"
+        class="p-2 rounded-lg bg-exo-dark-gray/80 border border-exo-yellow/30 text-exo-yellow hover:bg-exo-dark-gray hover:border-exo-yellow/50 cursor-pointer transition-colors"
+        onclick={handleClose}
+        title="Close"
+      >
+        <svg class="w-5 h-5" viewBox="0 0 24 24" fill="currentColor">
+          <path
+            d="M19 6.41L17.59 5 12 10.59 6.41 5 5 6.41 10.59 12 5 17.59 6.41 19 12 13.41 17.59 19 19 17.59 13.41 12 19 6.41z"
+          />
+        </svg>
+      </button>
+    </div>
+
+    <!-- svelte-ignore a11y_no_noninteractive_element_interactions, a11y_click_events_have_key_events -->
+    <img
+      {src}
+      alt=""
+      class="max-w-[90vw] max-h-[90vh] object-contain rounded-lg shadow-2xl"
+      transition:fly={{ y: 20, duration: 300, easing: cubicOut }}
+      onclick={(e) => e.stopPropagation()}
+    />
+  </div>
+{/if}

dashboard/src/lib/components/ImageParamsPanel.svelte

@@ -59,13 +59,14 @@
   }
 
   const sizeOptions: ImageGenerationParams["size"][] = [
+    "auto",
     "512x512",
     "768x768",
     "1024x1024",
     "1024x768",
     "768x1024",
-    "1024x1365",
-    "1365x1024",
+    "1024x1536",
+    "1536x1024",
   ];
 
   const qualityOptions: ImageGenerationParams["quality"][] = [
@@ -176,92 +177,90 @@
 <div class="border-b border-exo-medium-gray/30 px-3 py-2">
   <!-- Basic params row -->
   <div class="flex items-center gap-3 flex-wrap">
-    <!-- Size (hidden in edit mode - output size comes from input image) -->
-    {#if !isEditMode}
-      <div class="flex items-center gap-1.5">
-        <span class="text-xs text-exo-light-gray uppercase tracking-wider"
-          >SIZE:</span
+    <!-- Size -->
+    <div class="flex items-center gap-1.5">
+      <span class="text-xs text-exo-light-gray uppercase tracking-wider"
+        >SIZE:</span
+      >
+      <div class="relative">
+        <button
+          bind:this={sizeButtonRef}
+          type="button"
+          onclick={() => (isSizeDropdownOpen = !isSizeDropdownOpen)}
+          class="bg-exo-medium-gray/50 border border-exo-yellow/30 rounded pl-2 pr-6 py-1 text-xs font-mono text-exo-yellow cursor-pointer transition-all duration-200 hover:border-exo-yellow/50 focus:outline-none focus:border-exo-yellow/70 {isSizeDropdownOpen
+            ? 'border-exo-yellow/70'
+            : ''}"
         >
-        <div class="relative">
-          <button
-            bind:this={sizeButtonRef}
-            type="button"
-            onclick={() => (isSizeDropdownOpen = !isSizeDropdownOpen)}
-            class="bg-exo-medium-gray/50 border border-exo-yellow/30 rounded pl-2 pr-6 py-1 text-xs font-mono text-exo-yellow cursor-pointer transition-all duration-200 hover:border-exo-yellow/50 focus:outline-none focus:border-exo-yellow/70 {isSizeDropdownOpen
-              ? 'border-exo-yellow/70'
-              : ''}"
+          {params.size.toUpperCase()}
+        </button>
+        <div
+          class="absolute right-1.5 top-1/2 -translate-y-1/2 pointer-events-none transition-transform duration-200 {isSizeDropdownOpen
+            ? 'rotate-180'
+            : ''}"
+        >
+          <svg
+            class="w-3 h-3 text-exo-yellow/60"
+            fill="none"
+            viewBox="0 0 24 24"
+            stroke="currentColor"
           >
-            {params.size}
-          </button>
-          <div
-            class="absolute right-1.5 top-1/2 -translate-y-1/2 pointer-events-none transition-transform duration-200 {isSizeDropdownOpen
-              ? 'rotate-180'
-              : ''}"
-          >
-            <svg
-              class="w-3 h-3 text-exo-yellow/60"
-              fill="none"
-              viewBox="0 0 24 24"
-              stroke="currentColor"
-            >
-              <path
-                stroke-linecap="round"
-                stroke-linejoin="round"
-                stroke-width="2"
-                d="M19 9l-7 7-7-7"
-              />
-            </svg>
+            <path
+              stroke-linecap="round"
+              stroke-linejoin="round"
+              stroke-width="2"
+              d="M19 9l-7 7-7-7"
+            />
+          </svg>
+        </div>
+      </div>
+
+      {#if isSizeDropdownOpen}
+        <!-- Backdrop to close dropdown -->
+        <button
+          type="button"
+          class="fixed inset-0 z-[9998] cursor-default"
+          onclick={() => (isSizeDropdownOpen = false)}
+          aria-label="Close dropdown"
+        ></button>
+
+        <!-- Dropdown Panel - fixed positioning to escape overflow:hidden -->
+        <div
+          class="fixed bg-exo-dark-gray border border-exo-yellow/30 rounded shadow-lg shadow-black/50 z-[9999] max-h-48 overflow-y-auto overflow-x-hidden min-w-max"
+          style="bottom: calc(100vh - {sizeDropdownPosition()
+            .top}px + 4px); left: {sizeDropdownPosition().left}px;"
+        >
+          <div class="py-1">
+            {#each sizeOptions as size}
+              <button
+                type="button"
+                onclick={() => selectSize(size)}
+                class="w-full px-3 py-1.5 text-left text-xs font-mono tracking-wide transition-colors duration-100 flex items-center gap-2 {params.size ===
+                size
+                  ? 'bg-transparent text-exo-yellow'
+                  : 'text-exo-light-gray hover:text-exo-yellow'}"
+              >
+                {#if params.size === size}
+                  <svg
+                    class="w-3 h-3 flex-shrink-0"
+                    fill="currentColor"
+                    viewBox="0 0 20 20"
+                  >
+                    <path
+                      fill-rule="evenodd"
+                      d="M16.707 5.293a1 1 0 010 1.414l-8 8a1 1 0 01-1.414 0l-4-4a1 1 0 011.414-1.414L8 12.586l7.293-7.293a1 1 0 011.414 0z"
+                      clip-rule="evenodd"
+                    />
+                  </svg>
+                {:else}
+                  <span class="w-3"></span>
+                {/if}
+                <span>{size.toUpperCase()}</span>
+              </button>
+            {/each}
           </div>
         </div>
-
-        {#if isSizeDropdownOpen}
-          <!-- Backdrop to close dropdown -->
-          <button
-            type="button"
-            class="fixed inset-0 z-[9998] cursor-default"
-            onclick={() => (isSizeDropdownOpen = false)}
-            aria-label="Close dropdown"
-          ></button>
-
-          <!-- Dropdown Panel - fixed positioning to escape overflow:hidden -->
-          <div
-            class="fixed bg-exo-dark-gray border border-exo-yellow/30 rounded shadow-lg shadow-black/50 z-[9999] max-h-48 overflow-y-auto min-w-max"
-            style="bottom: calc(100vh - {sizeDropdownPosition()
-              .top}px + 4px); left: {sizeDropdownPosition().left}px;"
-          >
-            <div class="py-1">
-              {#each sizeOptions as size}
-                <button
-                  type="button"
-                  onclick={() => selectSize(size)}
-                  class="w-full px-3 py-1.5 text-left text-xs font-mono tracking-wide transition-colors duration-100 flex items-center gap-2 {params.size ===
-                  size
-                    ? 'bg-transparent text-exo-yellow'
-                    : 'text-exo-light-gray hover:text-exo-yellow'}"
-                >
-                  {#if params.size === size}
-                    <svg
-                      class="w-3 h-3 flex-shrink-0"
-                      fill="currentColor"
-                      viewBox="0 0 20 20"
-                    >
-                      <path
-                        fill-rule="evenodd"
-                        d="M16.707 5.293a1 1 0 010 1.414l-8 8a1 1 0 01-1.414 0l-4-4a1 1 0 011.414-1.414L8 12.586l7.293-7.293a1 1 0 011.414 0z"
-                        clip-rule="evenodd"
-                      />
-                    </svg>
-                  {:else}
-                    <span class="w-3"></span>
-                  {/if}
-                  <span>{size}</span>
-                </button>
-              {/each}
-            </div>
-          </div>
-        {/if}
-      </div>
-    {/if}
+      {/if}
+    </div>
 
     <!-- Quality -->
     <div class="flex items-center gap-1.5">
@@ -311,7 +310,7 @@
 
         <!-- Dropdown Panel - fixed positioning to escape overflow:hidden -->
         <div
-          class="fixed bg-exo-dark-gray border border-exo-yellow/30 rounded shadow-lg shadow-black/50 z-[9999] max-h-48 overflow-y-auto min-w-max"
+          class="fixed bg-exo-dark-gray border border-exo-yellow/30 rounded shadow-lg shadow-black/50 z-[9999] max-h-48 overflow-y-auto overflow-x-hidden min-w-max"
           style="bottom: calc(100vh - {qualityDropdownPosition()
             .top}px + 4px); left: {qualityDropdownPosition().left}px;"
         >

dashboard/src/lib/components/ModelCard.svelte

@@ -422,9 +422,16 @@
     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 });

dashboard/src/lib/components/ModelFilterPopover.svelte

@@ -24,6 +24,8 @@
     { id: "thinking", label: "Thinking" },
     { id: "code", label: "Code" },
     { id: "vision", label: "Vision" },
+    { id: "image_gen", label: "Image Gen" },
+    { id: "image_edit", label: "Image Edit" },
   ];
 
   // Size ranges
@@ -142,6 +144,45 @@
                   stroke-linejoin="round"
                 /><circle cx="12" cy="12" r="3" /></svg
               >
+            {:else if cap.id === "image_gen"}
+              <svg
+                class="w-3.5 h-3.5 inline-block"
+                viewBox="0 0 24 24"
+                fill="none"
+                stroke="currentColor"
+                stroke-width="1.5"
+                ><rect
+                  x="3"
+                  y="3"
+                  width="18"
+                  height="18"
+                  rx="2"
+                  ry="2"
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                /><circle cx="8.5" cy="8.5" r="1.5" /><path
+                  d="M21 15l-5-5L5 21"
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                /></svg
+              >
+            {:else if cap.id === "image_edit"}
+              <svg
+                class="w-3.5 h-3.5 inline-block"
+                viewBox="0 0 24 24"
+                fill="none"
+                stroke="currentColor"
+                stroke-width="1.5"
+                ><path
+                  d="M11 4H4a2 2 0 0 0-2 2v14a2 2 0 0 0 2 2h14a2 2 0 0 0 2-2v-7"
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                /><path
+                  d="M18.5 2.5a2.121 2.121 0 0 1 3 3L12 15l-4 1 1-4 9.5-9.5z"
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                /></svg
+              >
             {/if}
             <span class="ml-1">{cap.label}</span>
           </button>

dashboard/src/lib/components/ModelPickerGroup.svelte

@@ -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(
@@ -200,9 +240,42 @@
               stroke-width="1.5"
               title="Supports image generation"
             >
-              <rect x="3" y="3" width="18" height="18" rx="2" ry="2" />
+              <rect
+                x="3"
+                y="3"
+                width="18"
+                height="18"
+                rx="2"
+                ry="2"
+                stroke-linecap="round"
+                stroke-linejoin="round"
+              />
               <circle cx="8.5" cy="8.5" r="1.5" />
-              <path d="M21 15l-5-5L5 21" />
+              <path
+                d="M21 15l-5-5L5 21"
+                stroke-linecap="round"
+                stroke-linejoin="round"
+              />
+            </svg>
+          {:else if cap === "image_edit"}
+            <svg
+              class="w-3.5 h-3.5 text-white/40 flex-shrink-0"
+              viewBox="0 0 24 24"
+              fill="none"
+              stroke="currentColor"
+              stroke-width="1.5"
+              title="Supports image editing"
+            >
+              <path
+                d="M11 4H4a2 2 0 0 0-2 2v14a2 2 0 0 0 2 2h14a2 2 0 0 0 2-2v-7"
+                stroke-linecap="round"
+                stroke-linejoin="round"
+              />
+              <path
+                d="M18.5 2.5a2.121 2.121 0 0 1 3 3L12 15l-4 1 1-4 9.5-9.5z"
+                stroke-linecap="round"
+                stroke-linejoin="round"
+              />
             </svg>
           {/if}
         {/each}
@@ -211,7 +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}
@@ -222,13 +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
@@ -327,6 +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
@@ -351,7 +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>
 

dashboard/src/lib/components/ModelPickerModal.svelte

@@ -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,
@@ -363,6 +372,8 @@
       "deepseek",
       "gpt-oss",
       "llama",
+      "flux",
+      "qwen-image",
     ];
     return Array.from(families).sort((a, b) => {
       const aIdx = familyOrder.indexOf(a);
@@ -381,7 +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);
     }
 
@@ -425,13 +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) -
@@ -445,6 +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)) {
@@ -602,11 +669,12 @@
         families={uniqueFamilies}
         {selectedFamily}
         {hasFavorites}
+        hasRecents={hasRecentsTab}
         onSelect={(family) => (selectedFamily = family)}
       />
 
       <!-- Model list -->
-      <div class="flex-1 overflow-y-auto flex flex-col">
+      <div class="flex-1 overflow-y-auto scrollbar-hide flex flex-col">
         {#if selectedFamily === "huggingface"}
           <!-- HuggingFace Hub view -->
           <div class="flex-1 flex flex-col min-h-0">
@@ -624,7 +692,7 @@
             </div>
 
             <!-- Results list -->
-            <div class="flex-1 overflow-y-auto">
+            <div class="flex-1 overflow-y-auto scrollbar-hide">
               {#if hfIsLoadingTrending && hfTrendingModels.length === 0}
                 <div
                   class="flex items-center justify-center py-12 text-white/40"
@@ -709,6 +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"
@@ -740,6 +847,7 @@
               isFavorite={favorites.has(group.id)}
               {selectedModelId}
               {canModelFit}
+              {getModelFitStatus}
               onToggleExpand={() => toggleGroupExpanded(group.id)}
               onSelectModel={handleSelect}
               {onToggleFavorite}

dashboard/src/lib/components/PrefillProgressBar.svelte

@@ -0,0 +1,70 @@
+<script lang="ts">
+  import type { PrefillProgress } from "$lib/stores/app.svelte";
+
+  interface Props {
+    progress: PrefillProgress;
+    class?: string;
+  }
+
+  let { progress, class: className = "" }: Props = $props();
+
+  const percentage = $derived(
+    progress.total > 0
+      ? Math.round((progress.processed / progress.total) * 100)
+      : 0,
+  );
+
+  const etaText = $derived.by(() => {
+    if (progress.processed <= 0 || progress.total <= 0) return null;
+    const elapsedMs = performance.now() - progress.startedAt;
+    if (elapsedMs < 200) return null; // need a minimum sample window
+    const tokensPerMs = progress.processed / elapsedMs;
+    const remainingTokens = progress.total - progress.processed;
+    const remainingMs = remainingTokens / tokensPerMs;
+    const remainingSec = Math.ceil(remainingMs / 1000);
+    if (remainingSec <= 0) return null;
+    if (remainingSec < 60) return `~${remainingSec}s remaining`;
+    const mins = Math.floor(remainingSec / 60);
+    const secs = remainingSec % 60;
+    return `~${mins}m ${secs}s remaining`;
+  });
+
+  function formatTokenCount(count: number | undefined): string {
+    if (count == null) return "0";
+    if (count >= 1000) {
+      return `${(count / 1000).toFixed(1)}k`;
+    }
+    return count.toString();
+  }
+</script>
+
+<div class="prefill-progress {className}">
+  <div
+    class="flex items-center justify-between text-xs text-exo-light-gray mb-1"
+  >
+    <span>Processing prompt</span>
+    <span class="font-mono">
+      {formatTokenCount(progress.processed)} / {formatTokenCount(
+        progress.total,
+      )} tokens
+    </span>
+  </div>
+  <div class="h-1.5 bg-exo-black/60 rounded-full overflow-hidden">
+    <div
+      class="h-full bg-exo-yellow rounded-full transition-all duration-150 ease-out"
+      style="width: {percentage}%"
+    ></div>
+  </div>
+  <div
+    class="flex items-center justify-between text-xs text-exo-light-gray/70 mt-0.5 font-mono"
+  >
+    <span>{etaText ?? ""}</span>
+    <span>{percentage}%</span>
+  </div>
+</div>
+
+<style>
+  .prefill-progress {
+    width: 100%;
+  }
+</style>

dashboard/src/lib/components/TopologyGraph.svelte

@@ -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 ${edge.sinkRdmaIface || "?"}`;
+        missingIface = false;
+      } else {
+        ip = edge.sendBackIp || "?";
+        const ifaceInfo = getInterfaceLabel(edge.source, ip);
+        ifaceLabel = ifaceInfo.label;
+        missingIface = ifaceInfo.missing;
+      }
+
       entry.connections.push({
         from: edge.source,
         to: edge.target,
         ip,
-        ifaceLabel: ifaceInfo.label,
-        missingIface: ifaceInfo.missing,
+        ifaceLabel,
+        missingIface,
       });
       pairMap.set(key, entry);
     });
@@ -1120,15 +1137,17 @@
           .text(` (${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 +1155,46 @@
           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})` : ""}`,
+            );
         }
       }
     });

dashboard/src/lib/stores/app.svelte.ts

@@ -49,6 +49,7 @@ export interface NodeInfo {
   };
   last_macmon_update: number;
   friendly_name?: string;
+  os_version?: string;
 }
 
 export interface TopologyEdge {
@@ -56,6 +57,8 @@ export interface TopologyEdge {
   target: string;
   sendBackIp?: string;
   sendBackInterface?: string;
+  sourceRdmaIface?: string;
+  sinkRdmaIface?: string;
 }
 
 export interface TopologyData {
@@ -76,6 +79,8 @@ interface RawNodeIdentity {
   modelId?: string;
   chipId?: string;
   friendlyName?: string;
+  osVersion?: string;
+  osBuildVersion?: string;
 }
 
 interface RawMemoryUsage {
@@ -163,7 +168,7 @@ export interface ModelDownloadStatus {
 export interface PlacementPreview {
   model_id: string;
   sharding: "Pipeline" | "Tensor";
-  instance_meta: "MlxRing" | "MlxIbv" | "MlxJaccl";
+  instance_meta: "MlxRing" | "MlxJaccl";
   instance: unknown | null;
   memory_delta_by_node: Record<string, number> | null;
   error: string | null;
@@ -214,7 +219,6 @@ interface RawStateResponse {
     string,
     {
       MlxRingInstance?: Instance;
-      MlxIbvInstance?: Instance;
       MlxJacclInstance?: Instance;
     }
   >;
@@ -225,6 +229,19 @@ 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,
@@ -232,6 +249,25 @@ interface RawStateResponse {
   >;
   // Thunderbolt bridge cycles (nodes with bridge enabled forming loops)
   thunderboltBridgeCycles?: string[][];
+  // Disk usage per node
+  nodeDisk?: Record<
+    string,
+    { total: { inBytes: number }; available: { inBytes: number } }
+  >;
+  // MetaInstances (declarative instance constraints)
+  metaInstances?: Record<string, MetaInstanceData>;
+}
+
+export interface MetaInstanceData {
+  metaInstanceId: string;
+  modelId: string;
+  sharding: string;
+  instanceMeta: string;
+  minNodes: number;
+  nodeIds: string[] | null;
+  placementError: string | null;
+  consecutiveFailures: number;
+  lastFailureError: string | null;
 }
 
 export interface MessageAttachment {
@@ -255,6 +291,13 @@ export interface TokenData {
   topLogprobs: TopLogprob[];
 }
 
+export interface PrefillProgress {
+  processed: number;
+  total: number;
+  /** Timestamp (performance.now()) when prefill started. */
+  startedAt: number;
+}
+
 export interface Message {
   id: string;
   role: "user" | "assistant" | "system";
@@ -278,6 +321,7 @@ export interface Conversation {
   modelId: string | null;
   sharding: string | null;
   instanceType: string | null;
+  enableThinking: boolean | null;
 }
 
 const STORAGE_KEY = "exo-conversations";
@@ -287,13 +331,14 @@ const IMAGE_PARAMS_STORAGE_KEY = "exo-image-generation-params";
 export interface ImageGenerationParams {
   // Basic params
   size:
+    | "auto"
     | "512x512"
     | "768x768"
     | "1024x1024"
     | "1024x768"
     | "768x1024"
-    | "1024x1365"
-    | "1365x1024";
+    | "1024x1536"
+    | "1536x1024";
   quality: "low" | "medium" | "high";
   outputFormat: "png" | "jpeg";
   numImages: number;
@@ -317,7 +362,7 @@ export interface EditingImage {
 }
 
 const DEFAULT_IMAGE_PARAMS: ImageGenerationParams = {
-  size: "1024x1024",
+  size: "auto",
   quality: "medium",
   outputFormat: "png",
   numImages: 1,
@@ -425,6 +470,7 @@ function transformTopology(
       },
       last_macmon_update: Date.now() / 1000,
       friendly_name: identity?.friendlyName,
+      os_version: identity?.osVersion,
     };
   }
 
@@ -437,6 +483,8 @@ 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 +492,23 @@ 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,
+            });
           }
         }
       }
@@ -484,18 +545,43 @@ class AppStore {
   ttftMs = $state<number | null>(null); // Time to first token in ms
   tps = $state<number | null>(null); // Tokens per second
   totalTokens = $state<number>(0); // Total tokens in current response
+  prefillProgress = $state<PrefillProgress | null>(null);
+
+  // Abort controller for stopping generation
+  private currentAbortController: AbortController | null = null;
 
   // Topology state
   topologyData = $state<TopologyData | null>(null);
   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>>({});
+  metaInstances = $state<Record<string, MetaInstanceData>>({});
   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 +637,7 @@ class AppStore {
           modelId: conversation.modelId ?? null,
           sharding: conversation.sharding ?? null,
           instanceType: conversation.instanceType ?? null,
+          enableThinking: conversation.enableThinking ?? null,
         }));
       }
     } catch (error) {
@@ -740,6 +827,7 @@ class AppStore {
       modelId: derivedModelId,
       sharding: derivedSharding,
       instanceType: derivedInstanceType,
+      enableThinking: null,
     };
 
     this.conversations.unshift(conversation);
@@ -765,6 +853,7 @@ 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;
@@ -837,11 +926,7 @@ class AppStore {
 
     let instanceType: string | null = null;
     if (instanceTag === "MlxRingInstance") instanceType = "MLX Ring";
-    else if (
-      instanceTag === "MlxIbvInstance" ||
-      instanceTag === "MlxJacclInstance"
-    )
-      instanceType = "MLX RDMA";
+    else if (instanceTag === "MlxJacclInstance") instanceType = "MLX RDMA";
 
     let sharding: string | null = null;
     const inst = instance as {
@@ -1206,6 +1291,17 @@ 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 ?? {};
+      // MetaInstances
+      this.metaInstances = data.metaInstances ?? {};
       // Thunderbolt bridge cycles
       this.thunderboltBridgeCycles = data.thunderboltBridgeCycles ?? [];
       // Thunderbolt bridge status per node
@@ -1575,11 +1671,12 @@ class AppStore {
       if (!reader) throw new Error("No response body");
 
       let fullContent = prefixText;
+      let streamedThinking = "";
       const collectedTokens: TokenData[] = [...tokensToKeep];
 
       interface ChatCompletionChunk {
         choices?: Array<{
-          delta?: { content?: string };
+          delta?: { content?: string; reasoning_content?: string };
           logprobs?: {
             content?: Array<{
               token: string;
@@ -1600,6 +1697,7 @@ class AppStore {
         (parsed) => {
           const choice = parsed.choices?.[0];
           const delta = choice?.delta?.content;
+          const thinkingDelta = choice?.delta?.reasoning_content;
 
           // Collect logprobs data
           const logprobsContent = choice?.logprobs?.content;
@@ -1618,7 +1716,11 @@ class AppStore {
             }
           }
 
-          if (delta) {
+          if (thinkingDelta) {
+            streamedThinking += thinkingDelta;
+          }
+
+          if (delta || thinkingDelta) {
             if (firstTokenTime === null) {
               firstTokenTime = performance.now();
               this.ttftMs = firstTokenTime - requestStartTime;
@@ -1632,9 +1734,14 @@ class AppStore {
               this.tps = ((tokenCount - tokensToKeep.length) / elapsed) * 1000;
             }
 
-            fullContent += delta;
-            const { displayContent, thinkingContent } =
+            if (delta) {
+              fullContent += delta;
+            }
+            const { displayContent, thinkingContent: tagThinking } =
               this.stripThinkingTags(fullContent);
+            const combinedThinking = [streamedThinking, tagThinking]
+              .filter(Boolean)
+              .join("\n\n");
 
             if (this.activeConversationId === targetConversationId) {
               this.currentResponse = displayContent;
@@ -1646,7 +1753,7 @@ class AppStore {
               messageId,
               (m) => {
                 m.content = displayContent;
-                m.thinking = thinkingContent || undefined;
+                m.thinking = combinedThinking || undefined;
                 m.tokens = [...collectedTokens];
               },
             );
@@ -1658,11 +1765,14 @@ class AppStore {
 
       // Final update
       if (this.conversationExists(targetConversationId)) {
-        const { displayContent, thinkingContent } =
+        const { displayContent, thinkingContent: tagThinking } =
           this.stripThinkingTags(fullContent);
+        const finalThinking = [streamedThinking, tagThinking]
+          .filter(Boolean)
+          .join("\n\n");
         this.updateConversationMessage(targetConversationId, messageId, (m) => {
           m.content = displayContent;
-          m.thinking = thinkingContent || undefined;
+          m.thinking = finalThinking || undefined;
           m.tokens = [...collectedTokens];
           if (this.ttftMs !== null) m.ttftMs = this.ttftMs;
           if (this.tps !== null) m.tps = this.tps;
@@ -1770,11 +1880,12 @@ class AppStore {
       }
 
       let streamedContent = "";
+      let streamedThinking = "";
       const collectedTokens: TokenData[] = [];
 
       interface ChatCompletionChunk {
         choices?: Array<{
-          delta?: { content?: string };
+          delta?: { content?: string; reasoning_content?: string };
           logprobs?: {
             content?: Array<{
               token: string;
@@ -1795,6 +1906,7 @@ class AppStore {
         (parsed) => {
           const choice = parsed.choices?.[0];
           const delta = choice?.delta?.content;
+          const thinkingDelta = choice?.delta?.reasoning_content;
 
           // Collect logprobs data
           const logprobsContent = choice?.logprobs?.content;
@@ -1813,10 +1925,19 @@ class AppStore {
             }
           }
 
-          if (delta) {
-            streamedContent += delta;
-            const { displayContent, thinkingContent } =
+          if (thinkingDelta) {
+            streamedThinking += thinkingDelta;
+          }
+
+          if (delta || thinkingDelta) {
+            if (delta) {
+              streamedContent += delta;
+            }
+            const { displayContent, thinkingContent: tagThinking } =
               this.stripThinkingTags(streamedContent);
+            const combinedThinking = [streamedThinking, tagThinking]
+              .filter(Boolean)
+              .join("\n\n");
 
             // Only update currentResponse if target conversation is active
             if (this.activeConversationId === targetConversationId) {
@@ -1829,7 +1950,7 @@ class AppStore {
               assistantMessage.id,
               (msg) => {
                 msg.content = displayContent;
-                msg.thinking = thinkingContent || undefined;
+                msg.thinking = combinedThinking || undefined;
                 msg.tokens = [...collectedTokens];
               },
             );
@@ -1841,14 +1962,17 @@ class AppStore {
 
       // Final cleanup of the message (if conversation still exists)
       if (this.conversationExists(targetConversationId)) {
-        const { displayContent, thinkingContent } =
+        const { displayContent, thinkingContent: tagThinking } =
           this.stripThinkingTags(streamedContent);
+        const finalThinking = [streamedThinking, tagThinking]
+          .filter(Boolean)
+          .join("\n\n");
         this.updateConversationMessage(
           targetConversationId,
           assistantMessage.id,
           (msg) => {
             msg.content = displayContent;
-            msg.thinking = thinkingContent || undefined;
+            msg.thinking = finalThinking || undefined;
             msg.tokens = [...collectedTokens];
           },
         );
@@ -1869,6 +1993,11 @@ class AppStore {
     }
   }
 
+  /**
+   * Whether thinking is enabled for the current conversation
+   */
+  thinkingEnabled = $state(true);
+
   /**
    * Selected model for chat (can be set by the UI)
    */
@@ -1932,6 +2061,7 @@ class AppStore {
     reader: ReadableStreamDefaultReader<Uint8Array>,
     targetConversationId: string,
     onChunk: (parsed: T) => void,
+    onEvent?: Record<string, (data: unknown) => void>,
   ): Promise<void> {
     const decoder = new TextDecoder();
     let buffer = "";
@@ -1952,6 +2082,24 @@ class AppStore {
         const trimmed = line.trim();
         if (!trimmed) continue;
 
+        // Handle SSE comments (": key json") for prefill progress etc.
+        if (trimmed.startsWith(": ") && onEvent) {
+          const comment = trimmed.slice(2);
+          const spaceIdx = comment.indexOf(" ");
+          if (spaceIdx > 0) {
+            const key = comment.slice(0, spaceIdx);
+            if (onEvent[key]) {
+              try {
+                const parsed = JSON.parse(comment.slice(spaceIdx + 1));
+                onEvent[key](parsed);
+              } catch {
+                // Skip malformed JSON in comment
+              }
+            }
+          }
+          continue;
+        }
+
         if (trimmed.startsWith("data: ")) {
           const data = trimmed.slice(6);
           if (data === "[DONE]") continue;
@@ -2047,6 +2195,7 @@ class AppStore {
       textContent?: string;
       preview?: string;
     }[],
+    enableThinking?: boolean | null,
   ): Promise<void> {
     if ((!content.trim() && (!files || files.length === 0)) || this.isLoading)
       return;
@@ -2182,6 +2331,9 @@ class AppStore {
       let firstTokenTime: number | null = null;
       let tokenCount = 0;
 
+      const abortController = new AbortController();
+      this.currentAbortController = abortController;
+
       const response = await fetch("/v1/chat/completions", {
         method: "POST",
         headers: {
@@ -2194,7 +2346,11 @@ class AppStore {
           stream: true,
           logprobs: true,
           top_logprobs: 5,
+          ...(enableThinking != null && {
+            enable_thinking: enableThinking,
+          }),
         }),
+        signal: abortController.signal,
       });
 
       if (!response.ok) {
@@ -2208,10 +2364,11 @@ class AppStore {
       }
 
       let streamedContent = "";
+      let streamedThinking = "";
 
       interface ChatCompletionChunk {
         choices?: Array<{
-          delta?: { content?: string };
+          delta?: { content?: string; reasoning_content?: string };
           logprobs?: {
             content?: Array<{
               token: string;
@@ -2232,8 +2389,14 @@ class AppStore {
         reader,
         targetConversationId,
         (parsed) => {
+          // Clear prefill progress when first token data arrives
+          if (this.prefillProgress) {
+            this.prefillProgress = null;
+          }
+
           const choice = parsed.choices?.[0];
           const tokenContent = choice?.delta?.content;
+          const thinkingContent = choice?.delta?.reasoning_content;
 
           // Collect logprobs data
           const logprobsContent = choice?.logprobs?.content;
@@ -2252,7 +2415,11 @@ class AppStore {
             }
           }
 
-          if (tokenContent) {
+          if (thinkingContent) {
+            streamedThinking += thinkingContent;
+          }
+
+          if (tokenContent || thinkingContent) {
             // Track first token for TTFT
             if (firstTokenTime === null) {
               firstTokenTime = performance.now();
@@ -2269,11 +2436,16 @@ class AppStore {
               this.tps = (tokenCount / elapsed) * 1000;
             }
 
-            streamedContent += tokenContent;
+            if (tokenContent) {
+              streamedContent += tokenContent;
+            }
 
-            // Strip thinking tags for display and extract thinking content
-            const { displayContent, thinkingContent } =
+            // Use stripThinkingTags as fallback for any <think> tags still in content
+            const { displayContent, thinkingContent: tagThinking } =
               this.stripThinkingTags(streamedContent);
+            const combinedThinking = [streamedThinking, tagThinking]
+              .filter(Boolean)
+              .join("\n\n");
 
             // Only update currentResponse if target conversation is active
             if (this.activeConversationId === targetConversationId) {
@@ -2286,7 +2458,7 @@ class AppStore {
               assistantMessage.id,
               (msg) => {
                 msg.content = displayContent;
-                msg.thinking = thinkingContent || undefined;
+                msg.thinking = combinedThinking || undefined;
                 msg.tokens = [...collectedTokens];
               },
             );
@@ -2294,8 +2466,27 @@ class AppStore {
             this.persistConversation(targetConversationId);
           }
         },
+        {
+          prefill_progress: (data) => {
+            // TaggedModel wraps as {"PrefillProgressChunk": {...}}
+            // model_dump_json() uses snake_case (by_alias defaults to False)
+            const raw = data as Record<string, unknown>;
+            const inner = (raw["PrefillProgressChunk"] ?? raw) as {
+              processed_tokens: number;
+              total_tokens: number;
+            };
+            this.prefillProgress = {
+              processed: inner.processed_tokens,
+              total: inner.total_tokens,
+              startedAt: this.prefillProgress?.startedAt ?? performance.now(),
+            };
+          },
+        },
       );
 
+      // Clear prefill progress after stream ends
+      this.prefillProgress = null;
+
       // Calculate final TPS
       if (firstTokenTime !== null && tokenCount > 1) {
         const totalGenerationTime = performance.now() - firstTokenTime;
@@ -2304,14 +2495,17 @@ class AppStore {
 
       // Final cleanup of the message (if conversation still exists)
       if (this.conversationExists(targetConversationId)) {
-        const { displayContent, thinkingContent } =
+        const { displayContent, thinkingContent: tagThinking } =
           this.stripThinkingTags(streamedContent);
+        const finalThinking = [streamedThinking, tagThinking]
+          .filter(Boolean)
+          .join("\n\n");
         this.updateConversationMessage(
           targetConversationId,
           assistantMessage.id,
           (msg) => {
             msg.content = displayContent;
-            msg.thinking = thinkingContent || undefined;
+            msg.thinking = finalThinking || undefined;
             msg.tokens = [...collectedTokens];
             // Store performance metrics on the message
             if (this.ttftMs !== null) {
@@ -2326,20 +2520,31 @@ class AppStore {
         this.persistConversation(targetConversationId);
       }
     } catch (error) {
-      console.error("Error sending message:", error);
-      this.handleStreamingError(
-        error,
-        targetConversationId,
-        assistantMessage.id,
-        "Failed to get response",
-      );
+      if (error instanceof DOMException && error.name === "AbortError") {
+        // User stopped generation — not an error
+      } else {
+        console.error("Error sending message:", error);
+        this.handleStreamingError(
+          error,
+          targetConversationId,
+          assistantMessage.id,
+          "Failed to get response",
+        );
+      }
     } finally {
+      this.currentAbortController = null;
+      this.prefillProgress = null;
       this.isLoading = false;
       this.currentResponse = "";
       this.saveConversationsToStorage();
     }
   }
 
+  stopGeneration(): void {
+    this.currentAbortController?.abort();
+    this.currentAbortController = null;
+  }
+
   /**
    * Generate an image using the image generation API
    */
@@ -2852,6 +3057,18 @@ 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
    */
@@ -2954,21 +3171,26 @@ export const isLoading = () => appStore.isLoading;
 export const ttftMs = () => appStore.ttftMs;
 export const tps = () => appStore.tps;
 export const totalTokens = () => appStore.totalTokens;
+export const prefillProgress = () => appStore.prefillProgress;
 export const topologyData = () => appStore.topologyData;
 export const instances = () => appStore.instances;
+export const metaInstances = () => appStore.metaInstances;
 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();
 
 // Actions
+export const stopGeneration = () => appStore.stopGeneration();
 export const startChat = () => appStore.startChat();
 export const sendMessage = (
   content: string,
@@ -2979,7 +3201,8 @@ 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 +3245,8 @@ 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 +3263,12 @@ 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;
 

dashboard/src/lib/stores/recents.svelte.ts

@@ -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();

flake.nix

@@ -74,7 +74,6 @@
       perSystem =
         { config, self', inputs', pkgs, lib, system, ... }:
         let
-          fenixToolchain = inputs'.fenix.packages.complete;
           # Use pinned nixpkgs for swift-format (swift is broken on x86_64-linux in newer nixpkgs)
           pkgsSwift = import inputs.nixpkgs-swift { inherit system; };
         in
@@ -83,6 +82,9 @@
           _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";
@@ -112,7 +114,7 @@
           packages = lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin (
             let
               uvLock = builtins.fromTOML (builtins.readFile ./uv.lock);
-              mlxPackage = builtins.head (builtins.filter (p: p.name == "mlx") uvLock.package);
+              mlxPackage = builtins.head (builtins.filter (p: p.name == "mlx" && p.source ? git) uvLock.package);
               uvLockMlxVersion = mlxPackage.version;
             in
             {

nix/apple-sdk-overlay.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";
+  };
+}

nix/apple-sdk/metadata/versions.json

@@ -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="
+  }
+}

nix/mlx.nix

@@ -41,16 +41,16 @@ let
 
   mlx = stdenv.mkDerivation rec {
     pname = "mlx";
-    version = let v = "0.30.5"; in
+    version = let v = "0.30.7.dev20260220+13998a05"; 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;
 
     src = fetchFromGitHub {
-      owner = "ml-explore";
-      repo = "mlx";
-      tag = "v${version}";
-      hash = "sha256-OJk6jPlbaSlsUdk3ADz3tWcRzTWXRof3/q8Soe1AO6w=";
+      owner = "rltakashige";
+      repo = "mlx-jaccl-fix-small-recv";
+      rev = "13998a054715edcdc93618fb1496c79c7c25ff7c";
+      hash = "sha256-fAqA3hFwNBx7FcoGnhQsIFpAIRbC2EerACm4Fvne0Cc=";
     };
 
     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}")

pyproject.toml

@@ -17,9 +17,9 @@ 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-lm==0.30.6",
+    "mlx; sys_platform == 'darwin'",
+    "mlx[cpu]==0.30.6; sys_platform == 'linux'",
+    "mlx-lm==0.30.7",
     "tiktoken>=0.12.0", # required for kimi k2 tokenizer
     "hypercorn>=0.18.0",
     "openai-harmony>=0.0.8",
@@ -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]
@@ -57,10 +59,12 @@ dev = [
 [tool.uv.workspace]
 members = [
     "rust/exo_pyo3_bindings",
+    "bench",
 ]
 
 [tool.uv.sources]
 exo_pyo3_bindings = { workspace = true }
+mlx = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git", branch = "address-rdma-gpu-locks", marker = "sys_platform == 'darwin'" }
 #mlx-lm = { git = "https://github.com/davidmcc73/mlx-lm", branch = "stable" }
 # Uncomment to use local mlx/mlx-lm development versions:
 # mlx = { path = "/Users/Shared/mlx", editable=true }
@@ -129,7 +133,7 @@ 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",
 ]

python/parts.nix

@@ -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 +37,47 @@
 
       # 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;
+            mlx-cuda-13 = prev.mlx-cuda-13.overrideAttrs (old: {
+              buildInputs = (old.buildInputs or [ ]) ++ [
+                final.nvidia-cublas
+                final.nvidia-cuda-nvrtc
+                final.nvidia-cudnn-cu13
+                final.nvidia-nccl-cu13
+              ];
+              preFixup = ''
+                addAutoPatchelfSearchPath ${final.nvidia-cublas}
+                addAutoPatchelfSearchPath ${final.nvidia-cuda-nvrtc}
+                addAutoPatchelfSearchPath ${final.nvidia-cudnn-cu13}
+                addAutoPatchelfSearchPath ${final.nvidia-nccl-cu13}
+              '';
+              autoPatchelfIgnoreMissingDeps = [ "libcuda.so.1" ];
+            });
+            torch = ignoreMissing prev.torch;
+            triton = ignoreMissing prev.triton;
+          }
+        );
+
       pythonSet = (pkgs.callPackage inputs.pyproject-nix.build.packages {
         inherit python;
       }).overrideScope (
@@ -48,16 +86,28 @@
           overlay
           exoOverlay
           buildSystemsOverlay
+          linuxOverlay
         ]
       );
-      exoVenv = pythonSet.mkVirtualEnv "exo-env" workspace.deps.default;
+      # mlx-cpu and mlx-cuda-13 both ship mlx/ site-packages files; keep first.
+      # mlx-cpu/mlx-cuda-13 and nvidia-cudnn-cu12/cu13 ship overlapping files.
+      venvCollisionPaths = lib.optionals pkgs.stdenv.hostPlatform.isLinux [
+        "lib/python3.13/site-packages/mlx*"
+        "lib/python3.13/site-packages/nvidia*"
+      ];
+
+      exoVenv = (pythonSet.mkVirtualEnv "exo-env" workspace.deps.default).overrideAttrs {
+        venvIgnoreCollisions = venvCollisionPaths;
+      };
 
       # Virtual environment with dev dependencies for testing
-      testVenv = pythonSet.mkVirtualEnv "exo-test-env" (
+      testVenv = (pythonSet.mkVirtualEnv "exo-test-env" (
         workspace.deps.default // {
           exo = [ "dev" ]; # Include pytest, pytest-asyncio, pytest-env
         }
-      );
+      )).overrideAttrs {
+        venvIgnoreCollisions = venvCollisionPaths;
+      };
 
       mkPythonScript = name: path: pkgs.writeShellApplication {
         inherit name;
@@ -69,6 +119,16 @@
         text = ''exec python ${path} "$@"'';
       };
 
+      benchVenv = pythonSet.mkVirtualEnv "exo-bench-env" {
+        exo-bench = [ ];
+      };
+
+      mkBenchScript = name: path: pkgs.writeShellApplication {
+        inherit name;
+        runtimeInputs = [ benchVenv ];
+        text = ''exec python ${path} "$@"'';
+      };
+
       mkSimplePythonScript = name: path: pkgs.writeShellApplication {
         inherit name;
         runtimeInputs = [ pkgs.python313 ];
@@ -96,8 +156,9 @@
           exo = exoPackage;
           # Test environment for running pytest outside of Nix sandbox (needs GPU access)
           exo-test-env = testVenv;
-          exo-bench = mkPythonScript "exo-bench" (inputs.self + /bench/exo_bench.py);
         } // {
+        exo-bench = mkBenchScript "exo-bench" (inputs.self + /bench/exo_bench.py);
+        exo-eval-tool-calls = mkBenchScript "exo-eval-tool-calls" (inputs.self + /bench/eval_tool_calls.py);
         exo-get-all-models-on-cluster = mkSimplePythonScript "exo-get-all-models-on-cluster" (inputs.self + /tests/get_all_models_on_cluster.py);
       };
 
@@ -108,6 +169,21 @@
           ${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
+          '';
       };
     };
 }

resources/image_model_cards/exolabs--FLUX.1-Kontext-dev-4bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["ImageToImage"]
+family = "flux"
+quantization = "4bit"
+base_model = "FLUX.1 Kontext"
+capabilities = ["image_edit"]
 
 [storage_size]
 in_bytes = 15475325472

resources/image_model_cards/exolabs--FLUX.1-Kontext-dev-8bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["ImageToImage"]
+family = "flux"
+quantization = "8bit"
+base_model = "FLUX.1 Kontext"
+capabilities = ["image_edit"]
 
 [storage_size]
 in_bytes = 21426029632

resources/image_model_cards/exolabs--FLUX.1-Kontext-dev.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["ImageToImage"]
+family = "flux"
+quantization = ""
+base_model = "FLUX.1 Kontext"
+capabilities = ["image_edit"]
 
 [storage_size]
 in_bytes = 33327437952

resources/image_model_cards/exolabs--FLUX.1-Krea-dev-4bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = "4bit"
+base_model = "FLUX.1 Krea"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 15475325472

resources/image_model_cards/exolabs--FLUX.1-Krea-dev-8bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = "8bit"
+base_model = "FLUX.1 Krea"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 21426029632

resources/image_model_cards/exolabs--FLUX.1-Krea-dev.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = ""
+base_model = "FLUX.1 Krea"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 33327437952

resources/image_model_cards/exolabs--FLUX.1-dev-4bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = "4bit"
+base_model = "FLUX.1 Dev"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 15475325472

resources/image_model_cards/exolabs--FLUX.1-dev-8bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = "8bit"
+base_model = "FLUX.1 Dev"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 21426029632

resources/image_model_cards/exolabs--FLUX.1-dev.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = ""
+base_model = "FLUX.1 Dev"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 33327437952

resources/image_model_cards/exolabs--FLUX.1-schnell-4bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = "4bit"
+base_model = "FLUX.1 Schnell"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 15470210592

resources/image_model_cards/exolabs--FLUX.1-schnell-8bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = "8bit"
+base_model = "FLUX.1 Schnell"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 21415799872

resources/image_model_cards/exolabs--FLUX.1-schnell.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = ""
+base_model = "FLUX.1 Schnell"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 33306978432

resources/image_model_cards/exolabs--Qwen-Image-4bit.toml

@@ -4,6 +4,10 @@ hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
 uses_cfg = true
+family = "qwen-image"
+quantization = "4bit"
+base_model = "Qwen Image"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 26799533856

resources/image_model_cards/exolabs--Qwen-Image-8bit.toml

@@ -4,6 +4,10 @@ hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
 uses_cfg = true
+family = "qwen-image"
+quantization = "8bit"
+base_model = "Qwen Image"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 37014734400

resources/image_model_cards/exolabs--Qwen-Image-Edit-2509-4bit.toml

@@ -4,6 +4,10 @@ hidden_size = 1
 supports_tensor = false
 tasks = ["ImageToImage"]
 uses_cfg = true
+family = "qwen-image"
+quantization = "4bit"
+base_model = "Qwen Image Edit"
+capabilities = ["image_edit"]
 
 [storage_size]
 in_bytes = 26799533856

resources/image_model_cards/exolabs--Qwen-Image-Edit-2509-8bit.toml

@@ -4,6 +4,10 @@ hidden_size = 1
 supports_tensor = false
 tasks = ["ImageToImage"]
 uses_cfg = true
+family = "qwen-image"
+quantization = "8bit"
+base_model = "Qwen Image Edit"
+capabilities = ["image_edit"]
 
 [storage_size]
 in_bytes = 37014734400

resources/image_model_cards/exolabs--Qwen-Image-Edit-2509.toml

@@ -4,6 +4,10 @@ hidden_size = 1
 supports_tensor = false
 tasks = ["ImageToImage"]
 uses_cfg = true
+family = "qwen-image"
+quantization = ""
+base_model = "Qwen Image Edit"
+capabilities = ["image_edit"]
 
 [storage_size]
 in_bytes = 57445135488

resources/image_model_cards/exolabs--Qwen-Image.toml

@@ -4,6 +4,10 @@ hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
 uses_cfg = true
+family = "qwen-image"
+quantization = ""
+base_model = "Qwen Image"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 57445135488

resources/inference_model_cards/mlx-community--DeepSeek-V3.1-4bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "deepseek"
 quantization = "4bit"
 base_model = "DeepSeek V3.1"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 405874409472

resources/inference_model_cards/mlx-community--DeepSeek-V3.1-8bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "deepseek"
 quantization = "8bit"
 base_model = "DeepSeek V3.1"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 765577920512

resources/inference_model_cards/mlx-community--GLM-4.5-Air-8bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "8bit"
 base_model = "GLM 4.5 Air"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 122406567936

resources/inference_model_cards/mlx-community--GLM-4.5-Air-bf16.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "bf16"
 base_model = "GLM 4.5 Air"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 229780750336

resources/inference_model_cards/mlx-community--GLM-4.7-4bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "4bit"
 base_model = "GLM 4.7"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 198556925568

resources/inference_model_cards/mlx-community--GLM-4.7-6bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "6bit"
 base_model = "GLM 4.7"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 286737579648

resources/inference_model_cards/mlx-community--GLM-4.7-8bit-gs32.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "8bit"
 base_model = "GLM 4.7"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 396963397248

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-4bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "4bit"
 base_model = "GLM 4.7 Flash"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 19327352832

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-5bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "5bit"
 base_model = "GLM 4.7 Flash"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 22548578304

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-6bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "6bit"
 base_model = "GLM 4.7 Flash"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 26843545600

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-8bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "glm"
 quantization = "8bit"
 base_model = "GLM 4.7 Flash"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 34359738368

resources/inference_model_cards/mlx-community--GLM-5-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/GLM-5-8bit-MXFP8"
+n_layers = 78
+hidden_size = 6144
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "glm"
+quantization = "8bit"
+base_model = "GLM-5"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 790517400864

resources/inference_model_cards/mlx-community--GLM-5-MXFP4-Q8.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/GLM-5-MXFP4-Q8"
+n_layers = 78
+hidden_size = 6144
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "glm"
+quantization = "MXFP4-Q8"
+base_model = "GLM-5"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 405478939008

resources/inference_model_cards/mlx-community--GLM-5-bf16.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/GLM-5"
+n_layers = 78
+hidden_size = 6144
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "glm"
+quantization = "bf16"
+base_model = "GLM-5"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 1487822475264

resources/inference_model_cards/mlx-community--Kimi-K2-Thinking.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "kimi"
 quantization = ""
 base_model = "Kimi K2"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 706522120192

resources/inference_model_cards/mlx-community--Kimi-K2.5.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "kimi"
 quantization = ""
 base_model = "Kimi K2.5"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 662498705408

resources/inference_model_cards/mlx-community--MiniMax-M2.1-3bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "minimax"
 quantization = "3bit"
 base_model = "MiniMax M2.1"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 100086644736

resources/inference_model_cards/mlx-community--MiniMax-M2.1-8bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "minimax"
 quantization = "8bit"
 base_model = "MiniMax M2.1"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 242986745856

resources/inference_model_cards/mlx-community--MiniMax-M2.5-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/MiniMax-M2.5-4bit"
+n_layers = 62
+hidden_size = 3072
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "minimax"
+quantization = "4bit"
+base_model = "MiniMax M2.5"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 128666664960

resources/inference_model_cards/mlx-community--MiniMax-M2.5-6bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/MiniMax-M2.5-6bit"
+n_layers = 62
+hidden_size = 3072
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "minimax"
+quantization = "6bit"
+base_model = "MiniMax M2.5"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 185826705408

resources/inference_model_cards/mlx-community--MiniMax-M2.5-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/MiniMax-M2.5-8bit"
+n_layers = 62
+hidden_size = 3072
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "minimax"
+quantization = "8bit"
+base_model = "MiniMax M2.5"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 242986745856

resources/inference_model_cards/mlx-community--Qwen3-0.6B-4bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "4bit"
 base_model = "Qwen3 0.6B"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 342884352

resources/inference_model_cards/mlx-community--Qwen3-0.6B-8bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "8bit"
 base_model = "Qwen3 0.6B"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 698351616

resources/inference_model_cards/mlx-community--Qwen3-235B-A22B-Instruct-2507-4bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "4bit"
 base_model = "Qwen3 235B"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 141733920768

resources/inference_model_cards/mlx-community--Qwen3-235B-A22B-Instruct-2507-8bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "8bit"
 base_model = "Qwen3 235B"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 268435456000

resources/inference_model_cards/mlx-community--Qwen3-30B-A3B-4bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "4bit"
 base_model = "Qwen3 30B"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 17612931072

resources/inference_model_cards/mlx-community--Qwen3-30B-A3B-8bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "8bit"
 base_model = "Qwen3 30B"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 33279705088

resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-4bit.toml

@@ -3,6 +3,10 @@ n_layers = 48
 hidden_size = 2048
 supports_tensor = true
 tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "4bit"
+base_model = "Qwen3 Coder Next"
+capabilities = ["text", "code"]
 
 [storage_size]
 in_bytes = 45644286500

resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-5bit.toml

@@ -3,6 +3,10 @@ n_layers = 48
 hidden_size = 2048
 supports_tensor = true
 tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "5bit"
+base_model = "Qwen3 Coder Next"
+capabilities = ["text", "code"]
 
 [storage_size]
 in_bytes = 57657697020

resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-6bit.toml

@@ -3,6 +3,10 @@ n_layers = 48
 hidden_size = 2048
 supports_tensor = true
 tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "6bit"
+base_model = "Qwen3 Coder Next"
+capabilities = ["text", "code"]
 
 [storage_size]
 in_bytes = 68899327465

resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-8bit.toml

@@ -3,6 +3,10 @@ n_layers = 48
 hidden_size = 2048
 supports_tensor = true
 tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "8bit"
+base_model = "Qwen3 Coder Next"
+capabilities = ["text", "code"]
 
 [storage_size]
 in_bytes = 89357758772

resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-bf16.toml

@@ -3,6 +3,10 @@ n_layers = 48
 hidden_size = 2048
 supports_tensor = true
 tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "bf16"
+base_model = "Qwen3 Coder Next"
+capabilities = ["text", "code"]
 
 [storage_size]
 in_bytes = 157548627945

resources/inference_model_cards/mlx-community--Qwen3-Next-80B-A3B-Thinking-4bit.toml

@@ -6,7 +6,7 @@ tasks = ["TextGeneration"]
 family = "qwen"
 quantization = "4bit"
 base_model = "Qwen3 Next 80B"
-capabilities = ["text", "thinking"]
+capabilities = ["text", "thinking", "thinking_toggle"]
 
 [storage_size]
 in_bytes = 47080074240