app.py

import gradio as gr
from transformers import pipeline

pipe = pipeline("text-generation","TheBloke/Mistral-7B-Instruct-v0.1-GGUF")

def chat_fn(message, history):
    response = pipe(message, max_new_tokens=200)[0]["generated_text"]
    return response

gr.ChatInterface(fn=chat_fn).launch()