Update app.py

app.py

@@ -51,18 +51,18 @@ logging.getLogger("httpcore").setLevel(logging.WARNING)
 HF_TOKEN = os.environ.get("HF_TOKEN") or os.environ.get("HUGGING_FACE_HUB_TOKEN")
 os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")
 
-# Instruction-tuned Gemma, Qwen Coder, and R1 Distill 14B
+# Switched Code to NextCoder-14B for faster loading & lower VRAM overhead
 MODELS = {
     "Chat": "google/gemma-4-31B-it",
-    "Code": "Qwen/Qwen2.5-Coder-32B-Instruct",
+    "Code": "microsoft/NextCoder-14B",
     "Think": "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B",
 }
 
-# Generous ZeroGPU budgets so inference completes without server aborts
+# PRO tier gets the full 180s ceiling; Guest and Free tiers get standard allotments
 GPU_BUDGETS = {
-    "guest": {"Chat": 60, "Code": 55, "Think": 60},
-    "free": {"Chat": 75, "Code": 70, "Think": 75},
-    "pro": {"Chat": 100, "Code": 90, "Think": 120},
+    "guest": {"Chat": 45, "Code": 50, "Think": 50},
+    "free":  {"Chat": 60, "Code": 70, "Think": 70},
+    "pro":   {"Chat": 120, "Code": 180, "Think": 180},
 }
 
 THINK_OPEN = re.compile(r"<think>|◁think▷|<\|think\|>", re.I)
@@ -264,12 +264,12 @@ def history_to_prompt(history: list, mode: str, tokenizer) -> str:
         messages = [
             {
                 "role": "system",
-                "content": "Write correct, complete code. Prefer a working snippet over an essay.",
+                "content": "Write correct, clean, complete code. Prefer a working snippet over an essay.",
             },
             *visible,
         ]
     else:
-        # Chat: plain messages with no added system instructions or branding changes
+        # Chat: plain messages, pure natural responses with no added system instructions
         messages = visible
 
     if tokenizer.chat_template and messages:
@@ -383,7 +383,7 @@ def generate_response(
         streamer=streamer,
         max_new_tokens=max_new,
         do_sample=True,
-        temperature=0.6 if mode == "Think" else 0.4,
+        temperature=0.6 if mode == "Think" else 0.3,
         top_p=0.9,
         repetition_penalty=1.12,
         pad_token_id=tokenizer.pad_token_id,
@@ -393,7 +393,8 @@ def generate_response(
     worker.start()
 
     raw = ""
-    deadline = time.monotonic() + max(8, seconds - 4)
+    # Safe internal cutoff: 5 seconds before the ZeroGPU allocation expires
+    deadline = time.monotonic() + max(8, seconds - 5)
     try:
         for chunk in streamer:
             raw += chunk
@@ -422,7 +423,6 @@ def handle_user_submission(message, history):
         raise gr.Error("Enter a message first.")
     new_history = copy.deepcopy(history or [])
     new_history.append({"role": "user", "content": message})
-    # Instantly clear textbox and lock controls to avoid multiple clicks
     off = gr.update(interactive=False)
     return "", new_history, off, off, off
 
@@ -498,7 +498,7 @@ with gr.Blocks(
             choices=["Chat", "Code", "Think"],
             value="Chat",
             label="Mode",
-            info="Chat = Gemma 4 31B IT · Code = Qwen2.5-Coder-32B · Think = R1-Distill-Qwen-14B",
+            info="Chat = Gemma 4 31B IT · Code = NextCoder-14B · Think = R1-Distill-Qwen-14B",
         )
         chatbot = gr.Chatbot(
             label="Chat",