Jun Kim, creator of oMLX, joins Hugging Face to work on MLX
Hugging Face
- Open source
- Developer tools
- #omlx
- #mlx
- #transformers
Browse Open source topic page →
Hugging Face says Jun Kim, creator and maintainer of oMLX, has joined the company to support the MLX community. Per the post, oMLX stays Apache 2.0 under his leadership, and the aim is to give it stability and faster development as a funded project. Hugging Face also describes oMLX as a possible testbed for ideas it hopes to upstream, and names smoother movement from transformers model definitions to MLX implementations as one focus.
What Hugging Face announced
In a blog post dated September 22, 2026, Hugging Face announced that Jun Kim, described as the creator and maintainer of oMLX, has joined the company to support the MLX community. MLX is presented as Apple's framework for local AI, especially optimized for Apple Silicon. According to the post, the practical change for oMLX is funding: it moves from a side job to a fully maintained and funded project, which Hugging Face says should bring stability and hopefully faster development. The post states that oMLX stays Apache 2.0 and that Kim keeps leading it as before. Beyond oMLX, Hugging Face frames the project as a testbed for ideas, building on existing dependencies such as mlx-lm and mlx-vlm, with a stated wish to upstream work where it makes sense. One named focus is speeding the path from a transformers model definition to a reference MLX implementation usable by different engines. These are stated intentions, not reported completed work or measured results.
A scenario for a local-inference tool maintainer
A maintainer of an MLX-based engine or developer tool could treat this announcement as a signal to watch how oMLX evolves rather than as a reason to change dependencies now. Because the post says oMLX will act as a testbed and that Hugging Face wants to streamline getting transformers model definitions running on MLX, a team might pilot adopting a recently published transformers model on Apple Silicon through an MLX path and record where the workflow stalls. That scenario rests only on the post's stated goals; it assumes nothing about which engines will be supported or how quickly that transition will work in practice.
A small trial and how to judge it
A team could try a limited experiment: pick two or three recently released transformers models, follow the available MLX tooling to produce working implementations, and compare against their own baseline of doing the same task with current routes. Success could be judged by tasks completed, time spent, and where manual fixes were needed. Failure checks would include models that cannot be run at all or require changes the team cannot maintain. A maintainer should review any output before relying on it. Given that the post names the transition as a focus area but reports no delivered capability, the trial should be explicitly a probe of the current state, not an expectation of new features.
Trade-offs and when to hold back
The stated benefits are stability and direction from a funded maintainer, subject to the same Apache 2.0 license and leadership. The prerequisites are Apple Silicon hardware and the MLX ecosystem the post references. Unknowns include timing, which ideas would be upstreamed, and effects on other MLX inference engines and tools; a commenter raised that last point, and the post does not answer it. Commenters also reported the oMLX GitHub repository being unavailable shortly after publication, with speculation about automated moderation and a temporary fork, but the post itself says nothing about this, so the repository's status should be verified before depending on it. Teams needing guaranteed availability or long-term support commitments should wait for clear project status and maintainer statements rather than assume either from this announcement.
Context and limitations
- This draft is based on a short excerpt of the Hugging Face post; the full article and the linked projects were not analysed.
- The post describes intentions and hopes, such as faster development and upstreaming; it reports no completed milestones, benchmarks or independent tests.
- Commenters reported the oMLX GitHub repository was unavailable shortly after publication; the post does not address this and its status is unresolved here.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.