fix: heavy tier Dense→MoE→VLM

router/router.py

@@ -27,7 +27,7 @@ GPU_MAX_CONCURRENT = {
 # Context window sizes (tokens) — used for compaction signals
 GPU_CONTEXT = {
     "qwen3.6-35B-A3B": 131072,
-    "qwen3.6-27B-code": 65536,
+    "qwen3.6-27B-code": 98304,
     "qwen3.5-9b-vlm": 131072,
 }
 
@@ -190,8 +190,8 @@ def route(rd, tier):
     
     # TIER 3: Heavy reasoning — extremely large context or very long conversations
     if t > 50000 or turns > 25:
-        # Prefer models with larger context windows (MoE/VLM at 131K, Dense at 65K)
-        candidates = [m for m in ["qwen3.6-35B-A3B","qwen3.5-9b-vlm","qwen3.6-27B-code"] if m in avail]
+        # Dense first (98K, purpose-built for reasoning), then MoE/VLM 131K
+        candidates = [m for m in ["qwen3.6-27B-code","qwen3.6-35B-A3B","qwen3.5-9b-vlm"] if m in avail]
         result = select_best_gpu(candidates, "heavy_reasoning")
         if result: return result
     
@@ -308,10 +308,10 @@ def chat():
                     if raw: yield clean_unicode(raw)
             bcast()
             ctx_remaining = GPU_CONTEXT.get(model, 65536) - estimate_tokens(rd.get("messages",[]))
-            r = Response(stream_with_context(gen()), mimetype="text/event-stream")
-            r.headers["X-Context-Remaining"] = str(max(0, ctx_remaining))
-            r.headers["X-Context-Model"] = model
-            return r
+            sse_resp = Response(stream_with_context(gen()), mimetype="text/event-stream")
+            sse_resp.headers["X-Context-Remaining"] = str(max(0, ctx_remaining))
+            sse_resp.headers["X-Context-Model"] = model
+            return sse_resp
         data = clean_response(resp.json())
         for c in data.get("choices",[]):
             msg = c.get("message",{})