KernelBench Hub - Hard: leaderboard
KernelBench Hard evaluates autonomous coding agents on GPU kernel engineering tasks, measuring correctness and speed relative to hardware baselines.
Metric: Best % of Hardware Roofline. Source: kernelbench.com. Saturation forecast: Not forecast. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.8 Max | 69.26 | #35 |
| 2 | GLM-5.3 | 63.19 | #39 |
| 3 | Grok 4.6 | 62.17 | #26 |
| 4 | DeepSeek V4 Flash (0731) | 48.6 | #78 |
| 5 | Kimi K3 | 48.55 | #30 |
| 6 | DeepSeek V4.1 Flash (1017) | 45.86 | #20 |
| 7 | Claude Fable 5 | 43.03 | #16 |
| 8 | GPT-5.6 Sol | 38.71 | #19 |
| 9 | Claude Opus 5 | 37.08 | #12 |
| 10 | DeepSeek V4.1 Flash | 36.22 | #36 |
| 11 | GLM-5.3 Flash | 10.21 | #65 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=kernelbench-hub-hard · How It Works · Data refreshed daily, snapshot 2026-10-20.