MLX-VLM is a package for running and fine-tuning vision-language models on Apple silicon Macs using Apple’s MLX framework. It brings a wide range of popular VLMs, including LLaVA, PaliGemma, Pixtral, Molmo, Florence-2 and Idefics, under one consistent interface for local inference and training. The project has more than 5,600 stars on GitHub and was actively developed in 2026.
Every new vision-language model tends to ship with its own inference code, which makes comparing or combining them painful. MLX-VLM unifies them for Mac users: load a supported model, run image-and-text inference locally, generate with streaming output, or fine-tune with parameter-efficient adapters, all optimised for Apple silicon’s unified memory architecture. It is the vision counterpart to the popular MLX-LM tooling and a common starting point for multimodal experiments on macOS. New model architectures are typically added within days of their release, and the package handles the unglamorous details, such as image preprocessing, chat templates and tokenisation differences between model families, that otherwise make each VLM a bespoke integration project. For Apple-silicon owners it is often the fastest route from hearing about a new open vision-language model to actually running and adapting it on their own machine. Community examples cover tasks from image captioning to visual question answering.
Free. MLX-VLM is open source under the MIT license; you only need a Mac with Apple silicon and the models you want to run.
Mac developers and researchers who want to experiment with, evaluate or fine-tune multimodal models locally without cloud GPUs.