Hugging Face expands Apple local AI work

This briefing was generated with AI from the linked sources. Automated evidence checks are not human review. Results are source-reported unless explicitly stated otherwise. How this works.

Edition note. Originally prepared for September 22; published September 28 after approval. Source review was AI-assisted; runtime behavior and performance were not independently tested.

Transformers adds an Apple Silicon-focused GGUF path

Published September 22, 2026: Hugging Face announced support for loading GGUF models in Transformers with from_pretrained. The initial packed-inference work targets Apple Silicon and Qwen3.5, reusing ggml kernels through its kernels library. The announcement says users can also expose a loaded checkpoint through transformers serve with an OpenAI-compatible API.

Why it matters. This gives developers working in Python and PyTorch a route to experiment with GGUF checkpoints using standard Transformers workflows, including generation, evaluation, hooks, custom logits processors, and dequantized fine-tuning. It also potentially reduces the friction of using the same local model checkpoint with an OpenAI-compatible client interface.

Evidence. Transformers added GGUF support. (source) The first target is Apple Silicon local inference. (source)

Limits. This is a Hugging Face vendor report, not an independent benchmark. Its comparison used a MacBook Pro M2 Max and says the Transformers measurement included prefill while llama-bench reported decode-only throughput, so the conditions were not identical. The packed path is MPS-only, currently covers Qwen3.5 dense and MoE architectures plus compatible Qwen3.8 checkpoints, and padding and batching still need work. If compatible kernels are unavailable, the loader dequantizes and uses more memory.

Try this. Teams using Apple Silicon can test a GGUF model with the latest Transformers development version and compatible kernels, then evaluate the quantized model on their own task. The source recommends Q4_K_M as a starting point and says higher-precision variants may be tried where memory permits; quality tradeoffs depend on model and task.

Read the original · Published 2026-09-22 · Event 2026-09-22

oMLX creator Jun Kim joins Hugging Face

Published September 22, 2026: Hugging Face announced that Jun Kim, the creator and maintainer of oMLX, joined the company. According to the announcement, oMLX will remain Apache 2.0 and Kim will continue leading it.

Why it matters. The announcement positions the move as added maintenance capacity for oMLX and as part of Hugging Face’s local-AI work around MLX, Apple’s framework for local AI optimized for Apple Silicon. Hugging Face says a focus area is shortening the path from a Transformers model definition to a reference MLX implementation consumable by different inference engines.

Evidence. Hugging Face says oMLX remains under Apache 2.0. (source) The announcement says Kim continues as oMLX lead. (source)

Limits. The expected stability, development speed, upstreaming, and interoperability outcomes are plans and expectations reported by Hugging Face, rather than delivered results or independently verified project metrics. The source does not provide a release timeline, performance data, or a technical compatibility commitment for other engines.

Try this. Developers relying on oMLX should continue to track its project channels and license status, while treating future model-conversion and ecosystem integration work as prospective. The announcement suggests that work may matter most to teams building or maintaining MLX-based local inference tools.

Read the original · Published 2026-09-22 · Event 2026-09-22