Skill issue: stop deploying vision language models, use them with Skills — Merve Noyan, Hugging Face

Why watch

Merve Noyan demonstrates why developers should stop deploying Vision Language Models directly and instead use them as labeling tools to train specialized detectors like RF-DETR. Her toolkit uses a 9B parameter VLM for zero-shot labeling, then passes results to two smaller VLM judges that vote on minimum agreement rather than consensus to avoid losing too many examples. The entire workflow costs $3-4 on Hugging Face and produces detectors that outperform direct VLM use while running at 40fps on minimal hardware. She also warns about copyleft licensing in popular detection models and shows how to build vibe training pipelines with Apache 2.0 models.