CUDAHercules - Module-Level Mean Speed (Self-Refine): leaderboard
Metric: Mean expert-relative speed (the expert reference runtime divided by the generated code's runtime, averaged over the 43 architecture-general Class 2 module-level tasks, a task whose code fails the validator counting 0, so 1 matches the expert; self-refine@10: up to 10 revision rounds with compile and execution feedback (CUDAForge)). Source: arxiv.org. Saturation forecast: Around July 2027. 4 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 | 0.45 |
| 2 | GPT-5.4 | 0.3 |
| 3 | Qwen 3.5 122B A10B | 0 |
Interactive version: theaggregate.ai/benchmark?slug=cudahercules-module-level-mean-speed-self-refine · How It Works · Data refreshed daily, snapshot 2026-09-29.