What it is
A Python package built on Apple's MLX for running vision language models and omni models that take images, audio, and video on a Mac. It offers a CLI, a Python API, a Gradio chat UI, and an OpenAI-style FastAPI server with continuous batching, prefix caching, KV cache quantization, and speculative decoding. Developers and researchers use it to turn Hugging Face checkpoints into quantized MLX models, do OCR and multimodal chat offline, and fine-tune models on their own hardware.