Hugging Face brings oMLX creator in to build MLX support

What changes for oMLX

oMLX is expected to become more stable and develop faster now that Kim can work on it as a fully supported project rather than a side job.

Kim will continue to lead the project as before. Its Apache 2.0 license remains unchanged.

That continuity matters. Hugging Face is not replacing oMLX with an internal project; it is funding the maintainer who already directs it.

The infrastructure around MLX

Hugging Face has supported MLX since it became a “Christmas gift” from Avni and Angelos in 2023. The company sees the Hugging Face Hub as a place where users can find MLX models and publish their own.

Its broader goal is to help the community run local AI in any form, with the tools and building blocks needed to do so. oMLX is intended to serve as a testing ground for new ideas, built on dependencies including mlx-lm and mlx-vlm.

Hugging Face is already working with several projects:

The company also hopes to strengthen its relationships with Chen, Prince, Yagil and their teams to better serve the community.

The most concrete technical direction is a faster path from a model definition in transformers to a reference implementation for MLX. Different engines could then use that implementation while focusing on their own specialisations.

transformers has become a standard for describing machine-learning models. Hugging Face wants new transformers models to run on MLX with less work.

The bet behind the hire

I think the interesting part is not the staffing change by itself. It is the attempt to make MLX support a shared layer rather than a collection of isolated ports. If model definitions can move more quickly from transformers into a reference MLX implementation, libraries and engines do not each need to solve the same translation problem.

That makes oMLX more than a single local-inference project: it becomes a place where the ecosystem can test ideas against the foundations already provided by its dependencies. Hugging Face also plans to contribute useful work back to the main projects when appropriate, which could spread the benefits beyond oMLX.

The announcement is quiet about the operational details. It gives no funding figure and no delivery timetable for the reference implementations. My guess is that the success of this move will depend less on the hire than on whether those shared pieces arrive quickly enough for new models to reach MLX without another round of custom engineering.