Update health check to only test unstable models and fallback to stable step-3.7-flash

backend.py

@@ -467,17 +467,29 @@ def auth(authorization: str = None):
 
 async def check_models_health():
     global RECOMMENDED_MODEL
-    # Limit test to top 3 models to prevent NVIDIA NIM quota exhaustion
+    # Test only the unstable frontier models (the rest). 
+    # The stable ones (Step 3.7 Flash, Nemotron 3 Ultra, Qwen 2.5 Coder) are always free/working.
     models_to_test = [
-        "nvidia/nemotron-3-ultra-550b-a55b",
-        "stepfun-ai/step-3.7-flash",
-        "qwen/qwen2.5-coder-32b-instruct",
+        "moonshotai/kimi-k2.6",
+        "z-ai/glm-5.1",
+        "minimaxai/minimax-m3",
+        "minimaxai/minimax-m2.7",
+        "meta/llama-3.1-405b-instruct",
     ]
     
     best_model = None
     best_latency = 999.0
     
-    log_activity("Periodic health check started: verifying top 3 NIM models...")
+    # Mark stable models as permanently ONLINE in the status map
+    stable_models = [
+        "stepfun-ai/step-3.7-flash",
+        "nvidia/nemotron-3-ultra-550b-a55b",
+        "qwen/qwen2.5-coder-32b-instruct"
+    ]
+    for model in stable_models:
+        MODEL_STATUSES[model] = {"status": "ONLINE (Stable)", "latency": "Fast", "raw_latency": 0.1}
+
+    log_activity("Periodic health check started: verifying unstable frontier NIM models...")
     for model in models_to_test:
         start_time = time.time()
         try:
@@ -492,8 +504,7 @@ async def check_models_health():
             MODEL_STATUSES[model] = {"status": "ONLINE", "latency": f"{latency:.2f}s", "raw_latency": latency}
             log_activity(f"Model checked: {model} is ONLINE ({latency:.2f}s)")
             
-            # We want the model that is within 15 seconds
-            # and is the fastest (lowest latency)
+            # Choose the fastest online unstable model
             if latency < best_latency:
                 best_latency = latency
                 best_model = model
@@ -504,9 +515,11 @@ async def check_models_health():
             
     if best_model:
         RECOMMENDED_MODEL = best_model
-        log_activity(f"Best model selected: {RECOMMENDED_MODEL} ({best_latency:.2f}s)")
+        log_activity(f"Best frontier model selected: {RECOMMENDED_MODEL} ({best_latency:.2f}s)")
     else:
-        log_activity("Warning: All checked models failed or timed out!")
+        # Fallback to the stable Step 3.7 Flash if all frontier models are offline/throttled
+        RECOMMENDED_MODEL = "stepfun-ai/step-3.7-flash"
+        log_activity(f"All frontier models offline. Falling back to stable recommended model: {RECOMMENDED_MODEL}")
 
 async def periodic_health_check_loop():
     # Wait 10 seconds after startup before the first check to let the space boot fully