LLM GPU Kernel Generation: leaderboard
Trace-driven benchmark for LLM-generated GPU kernels (arXiv 2607.14541): the fraction of generated kernels that both compile and produce numerically correct results against reference execution traces.
Metric: Correctness Pass Rate (%). Source: arxiv.org. Saturation forecast: Around January 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 92 | #49 |
| 2 | GPT-5.5 | 91.1 | #31 |
| 3 | Qwen 3.7 Max | 84.8 | #75 |
| 4 | Kimi K2.6 | 81.5 | #104 |
| 5 | DeepSeek V4 Pro | 62.3 | #102 |
| 6 | GLM-5.1 | 46.2 | #97 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=llm-gpu-kernel-generation · How It Works · Data refreshed daily, snapshot 2026-10-20.