diff --git a/ESP32_ZACUS b/ESP32_ZACUS
index 8c076d8..82759ee 160000
--- a/ESP32_ZACUS
+++ b/ESP32_ZACUS
@@ -1 +1 @@
-Subproject commit 8c076d81d6efe4e7ddb6e21a4934e602275d19b0
+Subproject commit 82759ee536cd8694a1ad9d21808328dfdfe238af
diff --git a/tools/zacus-gateway/main.py b/tools/zacus-gateway/main.py
index b16d47a..f8bee9f 100644
--- a/tools/zacus-gateway/main.py
+++ b/tools/zacus-gateway/main.py
@@ -241,10 +241,13 @@ async def _chat_reply(http: httpx.AsyncClient, persona: str, history: list[dict]
     """
     messages = [{"role": "system", "content": persona + _PHONE_STYLE}] + history
     backends = [
-        # 70 s tolerates a ~47 s cold (re)load with margin; the keep-warm task
-        # keeps it ~2 s in normal operation. Total turn (STT ~10 s + this + say
-        # ~2 s) stays under the firmware's 90 s turn_client timeout.
-        ("local", settings.local_chat_url, settings.local_chat_model, 70.0),
+        # 8 s: a healthy warm granite answers in ~2 s, so 8 s gives margin while
+        # failing over FAST when the local server is cold, hung, or deadlocked.
+        # Waiting the full cold-load (~47 s) here would blow the turn budget —
+        # the player hears only the filler and the reply never lands in time.
+        # The keep-warm task is responsible for keeping granite warm; if it is
+        # unreachable we want the remote fallback immediately, not a 70 s stall.
+        ("local", settings.local_chat_url, settings.local_chat_model, 8.0),
         ("ailiance", settings.ailiance_tts_url, settings.ailiance_chat_model, 12.0),
     ]
     last_exc: Exception | None = None