A
argbe.tech - news
1min read

Claude-Built CUDA Kernel Skills, Now Portable to Open Models

Hugging Face introduced a workflow for turning a high-end model’s successful coding trace into a reusable “agent skill”, then evaluating how well that skill transfers to smaller open models for CUDA kernel work.

Hugging Face described how it uses a new tool, upskill, to turn an expert model’s agent run into a shareable skill and test whether smaller open models can follow it for CUDA kernel tasks.

  • The “teacher” setup uses Claude Opus 4.5 (via Claude Code) to build a CUDA kernel interactively, then exports the trace as the raw material for a skill.
  • The benchmark task focuses on writing CUDA kernels for diffusers, as a concrete stress test for domain-specific agent upskilling.
  • upskill can generate test cases from the trace and evaluate performance with the original trace versus applying the derived skill, highlighting when a skill helps—or increases token usage or hurts results on some models.
  • Skills are packaged as a directory with a SKILL.md file, and can be copied into common agent tool locations such as {agent}/skills/{skill_name}/SKILL.md (examples mentioned include codex, Cursor, and opencode).