Hugging Face Develops AI Skill for Converting Transformers Models to MLX Framework

Hugging Face has created an AI-powered Skill to help port language models from the transformers library to mlx-lm, aiming to make models available on MLX shortly after they land in transformers. The tool assists contributors by handling scaffolding tasks like finding model variants on the Hub, downloading checkpoints, and setting up editable installs. It also performs complex modeling tasks, paying attention to architecture details like RoPE configurations and verifying sensitive areas. For reviewers, the Skill produces PRs that follow mlx-lm conventions and include comprehensive reports with generation examples, numerical comparisons, and per-layer analyses. The approach addresses challenges posed by code agents, which since 2026 have increased PR volume tenfold while often missing implicit design contracts in codebases like transformers. The transformers library, with hundreds of contributors and over a billion downloads, prioritizes human-readable code with flat hierarchies. The Skill was developed using Claude Code and tested by porting models like GLM 4.7, incorporating learnings from contributors like @gabegoodhart. It includes a separate non-agentic test harness for reproducibility and transparency. Current limitations include handling of shared utilities in mlx-lm, vision-language models in mlx-vlm, and quantized model uploads. The project acknowledges Apple for open-sourcing MLX and the community for contributions.