mlx-vlm

Why we included this project

For anyone doing vision-language work on Apple silicon, this package removes most of the friction of getting modern models running locally. It provides inference and fine-tuning for a broad roster of models, including LLaVA, PaliGemma, Pixtral, MiniCPM-V, Gemma and several OCR-focused variants, exposed through a Python API, a CLI, and a FastAPI server. The practical extras are what stand out: speculative decoding, multi-image chat, KV-cache optimization, and per-model documentation with prompt formats and best practices, which is a real time-saver when you are porting a model to MLX. Teams that want private, on-device multimodal inference without cloud calls, or that need to benchmark models before committing, will find a solid base layer here.