Key Takeaways
- Hugging Face has hired Jun Kim, creator and maintainer of oMLX, to work full-time on the MLX community while the project stays under Apache 2.0.
- MLX has crossed 28,400 GitHub stars with roughly 4,800 pre-converted models in mlx-community, yet GGUF conversions still ship hours after new model releases.
- oMLX becomes an upstream testbed, shortening the path from transformers model definitions to native Apple Silicon inference and reducing a key maintainership risk.
Table of Contents
A Full-Time Home for the Maintainer Behind oMLX
Hugging Face has confirmed that Jun Kim, the creator and maintainer of oMLX, has joined its engineering team to support the MLX community on Apple Silicon.
The move, announced on September 22, 2026, turns oMLX from a side project into a funded, long-term maintainership role while the project remains under its Apache 2.0 license.
It also places the company more directly inside the fast-moving local AI stack built around Apple’s array framework.
What Changes for oMLX and the Broader MLX Ecosystem
The immediate effect is stability. According to Hugging Face’s announcement, instead of fitting development around other obligations, Kim will guide contributors and build for the long term without changing the project’s governance or license.
oMLX continues as an Apache 2.0 project, and Kim remains its lead maintainer.
oMLX stays Apache 2.0, and Jun keeps leading it as before.
The company expects oMLX to serve as a testbed for new modeling ideas while leaning on foundational dependencies such as mlx-lm and mlx-vlm.
That arrangement allows experimentation at the oMLX layer and lets improvements flow upstream where they benefit the wider MLX community.
MLX itself originated as Apple’s open framework for local AI, optimized specifically for Apple Silicon and first released in late 2023.
Since then it has become a major distribution target on the model-sharing platform, where the mlx-community repository now hosts roughly 4,800 pre-converted models.
MLX at 28,000 Stars: Performance Gains, Hardware Upside, and Format Tension
MLX has momentum. The GitHub repository has crossed 28,400 stars, 2,200 forks, and 262 contributors, with 2,295 commits as of September 2026.
Activity around MLX also grew faster than the broader open-model ecosystem during the first seven months of 2026, according to third-party analysis.
Independent performance tests have recorded generation speeds around 230 tokens per second on an M2 Ultra, while MLX has posted gains of 21 to 87 percent over llama.cpp on models under 14 billion parameters.
In the same evaluations, PyTorch MPS sat at roughly 7 to 9 tokens per second, making MLX the more practical Apple Silicon option for many local inference workloads.
Those figures come with caveats. Serving maturity still trails llama.cpp, and long-context prefill can degrade beyond 8,000 tokens without Flash Attention.
Core constrained decoding is also absent, which pushes some production teams toward llama.cpp for specific serving requirements.
- Fine-tuning: LoRA, QLoRA, and DoRA adapters run natively on Apple Silicon.
- Quantization: The framework supports 2-, 3-, 4-, 6-, and 8-bit formats for memory-constrained inference.
- Serving: An OpenAI-compatible server is available through mlx_lm.server, and Ollama 0.19 and later can use an MLX backend.
- Scale: Distributed inference can run across multiple Macs.
The competitive tension with GGUF remains real. GGUF is a single-file format from the llama.cpp ecosystem, while MLX models are folders containing safetensors weights, config files, and tokenizers.
GGUF conversions often appear within hours of a new model release, while MLX conversions can take days or weeks, but MLX tends to deliver faster generation once the model is loaded.
On M1 and M2 chips, bf16 is not hardware-accelerated, so converting MLX weights to fp16 can recover prefill speed; M3 and newer chips run bf16 natively.
GGUF remains inference-only, whereas MLX supports fine-tuning on-device, which makes the oMLX testbed role specifically valuable for model implementation work.
Apple’s August hardware announcements, including larger unified memory pools and neural engines tuned for agentic workloads, increase the upside for local AI infrastructure.
Awni Hannun, an MLX co-creator, reportedly left Apple earlier in 2026, putting more weight on external maintainership and community-driven development.
From Transformers Definitions to Native Apple Silicon Inference
The deeper play is workflow acceleration. The company wants to shorten the path from a transformers model definition to a reference MLX implementation that other engines can consume.
If that pipeline works, each engine can concentrate on its unique features instead of re-implementing model logic from scratch.
That matters now because new transformers models often appear as GGUF within hours, while MLX conversions lag by days or weeks; reducing that gap could materially change local AI adoption on Apple hardware.
For developers building on Apple Silicon, the hiring removes a key maintainership risk at the moment MLX adoption is accelerating, and it turns oMLX into an upstream testbed where modeling ideas can move faster from definition to production inference. For teams building AI automation and content pipelines around fast-moving model ecosystems, programmatic SEO and AI automation is how Andres SEO Expert approaches scalable search growth — contact us here.
Frequently Asked Questions
What is oMLX and who maintains it?
oMLX is an open-source project created and led by Jun Kim that builds on top of Apple Silicon inference libraries such as mlx-lm and mlx-vlm. It remains an Apache 2.0 project, and Kim continues as its lead maintainer after joining Hugging Face in September 2026.
What does Hugging Face hiring the oMLX maintainer change?
The main change is long-term stability, not governance. Jun Kim now works on oMLX full time instead of around other obligations, so he can guide contributors and develop the project for the long term. The license stays Apache 2.0, the project structure is unchanged, and oMLX is expected to act as a testbed whose improvements flow upstream to the wider MLX community.
Is MLX faster than llama.cpp on Apple Silicon?
For generation speed on many local models, yes. Independent tests recorded around 230 tokens per second on an M2 Ultra, and MLX showed gains of 21 to 87 percent over llama.cpp on models under 14 billion parameters. PyTorch MPS measured only about 7 to 9 tokens per second in the same evaluations. Caveats apply: llama.cpp still has more mature serving features, and long-context prefill on MLX can degrade beyond 8,000 tokens without Flash Attention.
What is the difference between MLX and GGUF model formats?
GGUF is a single-file format from the llama.cpp ecosystem, while MLX models are folders containing safetensors weights, config files, and tokenizers. GGUF conversions often appear within hours of a model release, whereas MLX conversions can take days or weeks, but MLX usually generates faster once the model is loaded. GGUF is inference-only, while MLX also supports on-device fine-tuning.
Does MLX support fine-tuning and quantization on Apple Silicon?
Yes. MLX runs LoRA, QLoRA, and DoRA adapters natively on Apple Silicon, which GGUF does not support. It also handles 2-, 3-, 4-, 6-, and 8-bit quantization formats for memory-constrained inference.
Can MLX serve models or scale across multiple Macs?
MLX offers an OpenAI-compatible server through mlx_lm.server, and Ollama 0.19 and later can use an MLX backend. Distributed inference can also run across multiple Macs, although serving maturity still trails llama.cpp for some production requirements.
Why is MLX conversion slower than GGUF for new models?
New transformers models typically appear as GGUF within hours, while MLX conversions lag by days or weeks. Hugging Face wants to shorten the path from a transformers model definition to a reference MLX implementation that other engines can consume, so each engine can focus on unique features instead of re-implementing model logic from scratch. Closing that gap could materially accelerate local AI adoption on Apple hardware.
