mizorewww/laya-mlx ? reverse-engineered prompt

Build me a small Python project that runs typed decision models locally on Apple Silicon with MLX. I want to load a pretrained checkpoint, send in some state plus a question, and get back a clear answer like a choice, a score, or a true false style result, without generating normal text. It should work fully offline after the first download, and it should feel fast on a Mac.

Please include a simple Python API, a command line demo that can play a little Snake game by making a decision each move, and a couple of examples showing how to call it with English and multilingual inputs. Keep it focused on inference only, no training, no cloud calls, and no PyTorch. If there are any current MLX or Hugging Face details that matter, look them up online and use the right approach. Make the README and examples easy to follow so someone can install it and try it right away.