CUDAHercules - Module-Level Pass Rate (Self-Refine): leaderboard
Metric: Pass rate (%; share of the 43 architecture-general Class 2 module-level tasks whose generated CUDA passes the task validator; self-refine@10: up to 10 revision rounds with compile and execution feedback (CUDAForge)). Source: arxiv.org. Saturation forecast: Around December 2026. 4 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 | 67 |
| 2 | GPT-5.4 | 56 |
| 3 | Qwen 3.5 122B A10B | 2 |
Interactive version: theaggregate.ai/benchmark?slug=cudahercules-module-level-pass-rate-self-refine · How It Works · Data refreshed daily, snapshot 2026-09-26.