The problem is even small local models tend to be rather demanding so trying to both render a game and run an LLM is going to be extremely taxing