Hugging Face introduces a Skill and test harness to help port language models from the transformers library to mlx-lm using AI coding agents. The post addresses the growing problem of low-quality agent-generated PRs flooding open source projects, and explains how this tool is designed to assist contributors and reviewers rather than automate contributions blindly. The Skill guides Claude Code through the full porting process — discovering models, writing MLX implementations, running per-layer comparisons, and verifying RoPE configs and dtypes — while producing PRs that follow mlx-lm conventions. A separate non-agentic test harness provides reproducible verification independent of LLM hallucinations. The post also reflects on how meaningful open source contribution must still involve human ownership and engagement with reviewer feedback.
Table of contents
TL;DRThe advent of code agentsWhat does this have to do with MLX?What we didHow we did itTest harnessHow to use the SkillNext steps and known shortcomingsConclusionResourcesThank you!Sort: