mlx_KM

Route Apple Silicon MLX work to the right specialist workflow. Use this whenever the user says MLX, mlx-lm, mlx-vlm, mlx-embeddings, Apple Silicon model conversion, local Mac model inference, quantiza

SKILL.md

GitHub

MLX router

Use this as the first stop for Apple Silicon MLX work. Keep it short: its job is to choose the right specialist skill, not to duplicate all MLX procedures.

Routing

Start with mlx-model-repo-audit when the user gives a Hugging Face repo, local model folder, fine-tuned checkpoint, or asks whether a model is MLX-compatible. Auditing first prevents the common mistake of routing multimodal models through text-only LLM tooling.

Use mlx-llm-conversion when the model is a text-only causal/chat LLM and the task is conversion, loading, generation, LoRA/QLoRA/fusion, or mlx-lm usage.

Use mlx-vlm-conversion when the model consumes images, audio, video, PDFs, screenshots, OCR inputs, scientific figures, or any multimodal processor artifacts. VLM behavior depends on processor files and media token handling, not just weights.

Use mlx-embedding-reranker when the output is vectors, similarity scores, retrieval ranking, or document/image embedding rather than generated text.

Use mlx-quantization-quality when the user asks about bit width, memory savings, speed/quality tradeoffs, mxfp/nvfp modes, group size, or whether a quantized model is good enough.

Use mlx-local-benchmarking when a model already runs and the user wants cold-load time, first-token latency, throughput, peak memory, batch behavior, or variant comparisons on the Mac Studio/MacBook.

Use mlx-troubleshooting-repro when there is a concrete error, crash, bad output, missing keys, processor mismatch, OOM, package-version issue, or GitHub issue to file.

Use mlx-hub-packaging only after validation passes and the user wants a reusable local package, Hugging Face upload, model card, or MLX Community-style repo.

Use mlx-speculative-decoding when the user asks about speculative decoding, draft models, DFlash, MTP drafters, faster generation through speculation, or --draft-model for mlx-lm/mlx-vlm. The skill captures runtime support gaps, decision thresholds, drafter selection, and verified per-chip results so the agent doesn't repeat known-bad experiments.

Default sequence

For an unknown model, use this sequence:

  1. Audit the repo/model.
  2. Choose LLM, VLM, or embedding/reranker path.
  3. Convert or load a baseline first, ideally unquantized when feasible.
  4. Quantize only after baseline behavior is known.
  5. Benchmark once the model runs.
  6. Package or upload only after validation evidence exists.

References

Read references/recommended-suite.md for the full suite map, references/research-brief.md for the ecosystem snapshot, and references/validation-checklists.md before claiming a conversion is complete.