Been testing out GPT-4-x-Alpaca 13b with oobabooga UI. Fits in 11 gigs of VRAM locally, can alternatively load on CPU, has reasonable 3-5 second response times (GPU tested), exposes a local API for use, and handles prompts similarly to ChatGPT. Might be a nice option in the future, if we could somehow point the game to a different endpoint API.