← Back to the wire

Custom Kernels for All from Codex and Claude

AchievementModelFeb 13, 2026

Hugging Face researchers built an agent skill that teaches coding agents to write production CUDA kernels. Ben Burtenshaw, Sayak Paul, Aritra Roy Gosthipaty, and Shaun Smith directed Claude and Codex at two targets: LTX-Video in diffusers and Qwen3-8B in transformers. The agents produced working kernels with correct PyTorch bindings and benchmarks on H100 hardware, filling the gap between kernel distribution and kernel creation.

Receipt № 5571 source · awaiting confirmation ◐

Evidence

1source· awaiting independent confirmation

No score is assigned. Sources and their independence are shown in the citation chain below.

Citation chain · 1 source

Qwen3-8BModelCodexModelClaudeModelBen BurtenshawPersonSayak PaulPersonShaun SmithPersonLTX-VideoModelAritra Roy GosthipatyPersonHugging FaceCompany
Canonical: https://huggingface.co/blog/custom-cuda-kernels-agent-skills