[Tepix]

A relatively small and fast LLM (30b MoE) would probably suffice. It‘s something you can run locally with a nice GPU.