CUDABeaver - Pass@1: leaderboard

CUDABeaver (2026) gives a model a real failing CUDA kernel from an LLM generation run, its build and test commands and the raw error, and asks for a repair that stays correct without falling back to a slower program. This board is the percentage of the 213 tasks solved by the first repair attempt, counting a repair only if it runs at least 0.7 times as fast as the reference; higher is better.

Metric: Share of the 213 broken-start CUDA debugging tasks (%) whose first repair attempt compiles, passes the task tests and runs at least 0.7 times as fast as the reference (performance gate p=0.7), default protocol: category-aware error feedback (level L3), iterative sampling at temperature 0.7; higher is better. Source: arxiv.org. Saturation forecast: Around May 2028. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3.6 Plus15.5#245
2Gemma 4 31B (IT)14.7#354
3GPT-5.414.5#138
4Kimi K2.613.6#197
5Qwen 3.6 27B10.3#358
6GLM-4.77.5#383

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=cudabeaver-pass-1 · How It Works · Data refreshed daily, snapshot 2026-11-22.