What problem does it solve? Determining which locally cached Hugging Face models are actually runnable vision-language models is error-prone: the cache contains text-only, embedding, reranking, drafter, and image-generation repos that pass naive file checks. This Skill explains the two-layer discovery contract (cache-layout filter plus image-capability classification) used by check_models so you can list, audit, and reason about eligible models without reimplementing cache-scan logic. ## Core Features & Use Cases - Cache-layout filtering: Applies the mlx-vlm server-style rule (model repo type, main revision, config.json, tokenizer_config.json, safetensors weights) to identify structurally valid cached repos. - Image-capability classification: Tri-state verdicts (yes/no/unknown) from bounded config metadata, distinguishing VLMs from drafters, embeddings, rerankers, and generation models, with evidence recorded in results.jsonl. - Architecture and hub pre-checks: Mirrors the upstream --check-arch tier against installed mlx_vlm model packages, and judges hub checkpoints before download via tools.hub_precheck. - Use Case: Run python -m check_models --dry-run to see exactly which cached models default discovery would benchmark, with skip reasons for excluded repos. ## Quick Start Ask the assistant to list the Hugging Face cached models that check_models default discovery would select, including skip reasons for any excluded repos.
