We Got Claude to Build CUDA Kernels and teach open models!

Source: HuggingFace Date: 2026-01-28 URL: https://huggingface.co/blog/upskill

Summary

Library release and tutorial: upskill — HF tool for generating structured agent skills (task-specific instruction packages ~500 tokens) using Claude Opus 4.5, then evaluating knowledge transfer to smaller/cheaper models. Demonstrated with CUDA kernel generation: Claude Opus baseline 60% → 95% with skill (+35%); local GLM-4.7-Flash-GGUF 40% → 85% with skill (+45%). CLI-based: upskill generate, upskill eval with support for local GGUF models via OpenAI-compatible endpoints.

Implications

Open-weights ecosystem health. The “Robin Hood” framing — use expensive powerful models to create skill packages, deploy with cheap/local models — is a legitimate inference cost reduction strategy for repeated domain-specific tasks. The 45-point improvement for a local GGUF model using a skill generated by Claude Opus demonstrates that knowledge transfer via structured prompt engineering is quantitatively real, not just a qualitative improvement.

Model release cadence (agent reasoning). upskill positions Claude Opus as a skill-generation oracle whose outputs improve open-weights model performance on specific tasks. This is a distinct use case from general chat or reasoning — using frontier models as knowledge extractors whose outputs can be packaged for distribution. If the skill generation pattern generalizes, it creates a new category of open-weights-compatible capability enhancement that doesn’t require fine-tuning.