Update app.py

app.py

@@ -3,12 +3,13 @@ import gradio as gr
 model = gr.load("models/mistralai/Mixtral-8x7B-Instruct-v0.1")
 
 def chat(prompt, response_type):
+    response = model(prompt)
+    max_tokens = 200
     if response_type == "Short":
-        return model(prompt, max_tokens=50)
+        max_tokens = 50
     elif response_type == "Medium":
-        return model(prompt, max_tokens=100)
-    else:
-        return model(prompt, max_tokens=200)
+        max_tokens = 100
+    return response[:max_tokens]
 
 demo = gr.Interface(
     fn=chat,