KernelZero-7B co-evolution beats Claude 4.5 Sonnet on CUDA kernel generation
Changxin Ke · hf · 2026-09-29
KernelZero is a co-evolution framework with a Proposer that generates Torch modules from API sets and a Coder that translates them into CUDA/Triton kernels, tackling training data scarcity and the correctness-performance tradeoff.
Key points:
- A frontier-driven module generation mechanism continuously produces capability-aligned training modules targeting the Coder's current weaknesses;
- Correctness-Aware GRPO (CA-GRPO) optimizes performance only after correctness is reliable, forming an automatic curriculum;
- KernelZero-7B surpasses Claude-4.5-Sonnet on CUDA and DeepSeek-V4-Pro on Triton: KernelBench Level 1/2 CUDA pass@1 of 75.8%/69.6% with pass@10 of 100%/97%; Triton pass@1 of 77.2%/72.5%.
More from coding & agent
- Claude Code user wants a default-model setting for workflow subagents after runaway 200-agent session — Liu_eroteme · 2026-09-29
- When Shopify WebMCP checkout disagrees with the visible UI, who should the agent trust? — daani_maas · 2026-09-29
- Stripe and Notion converge on the same bet: agents will be both sides of knowledge work — ivanhzhao · 2026-09-29
- Seth Rosen pitches structured agent memory over markdown — and Notion's CEO coyly responds — ivanhzhao · 2026-09-29
- Dev open-sources GitHub Radar: one page tracking all issues and PRs across orgs and repos — dSebastien · 2026-09-29
- Model-suggested performance optimizations remain 'laughably bad' even with profiler guidance — remilouf · 2026-09-29