Update app.py

app.py

@@ -27,8 +27,8 @@ def load_model():
     # Load the locally downloaded model here, specifying the device
     llm = CTransformers(
         model="TheBloke/zephyr-7B-beta-GGUF",
-        model_type="llama",
-        max_new_tokens=1096,
+        model_type="mistral",
+        max_new_tokens=4384,
         temperature=0.2,
         repetition_penalty=1.13,
         device=device  # Set the device explicitly during model initialization