KernelGenBench-MS (Pass@1): leaderboard

Metric: Clean pass rate (%; share of the 210 KernelGenBench-MS operators (110 PyTorch ATen, 50 vLLM, 50 cuBLAS) whose generated Triton kernel passes numerical validation on every test case of its combinatorial shape, dtype and layout suite and every anti-hack check, as a drop-in replacement; direct zero-shot generation, one sample at temperature 0, no execution feedback; NVIDIA A100; 30-minute budget per operator). Source: arxiv.org. Saturation forecast: Around December 2026. 4 models tracked.

Top models

#ModelScore
1Claude Opus 4.641
2GLM-521
3Qwen 3.5 27B7
4MiniMax-M2.52

Interactive version: theaggregate.ai/benchmark?slug=kernelgenbench-ms-pass-1 · How It Works · Data refreshed daily, snapshot 2026-09-29.