diff --git a/backend/src/spicebook/chat/llm.py b/backend/src/spicebook/chat/llm.py
index 12463c4..322e481 100644
--- a/backend/src/spicebook/chat/llm.py
+++ b/backend/src/spicebook/chat/llm.py
@@ -97,7 +97,10 @@ async def chat_completion_stream(
             "messages": messages,
             "temperature": 0.3,
             "max_tokens": settings.chat_max_tokens,
-            "num_ctx": settings.chat_max_tokens,
+            # No num_ctx: it makes Ollama reload the model at a different context
+            # size, which is an unload+load. That stalls whenever the backend's
+            # eviction path is unhealthy, and it would *shrink* the context the
+            # resident model already serves. Let the gateway pick the context.
             "stream": True,
         },
     ) as resp: