KernelBench Hard — B200 Mean Peak Fraction
Code · 2026-04-29
KernelBench Hard mean peak fraction on the six-problem B200 board. The score is the arithmetic mean of hardware-roofline percentages across cells that KernelBench.com marks valid; failed or audit-flagged cells do not enter that performance mean.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 13.6 |
| GLM-5.2 | 9.7 |
| Opus 4.8 | 9.6 |
| Kimi K3 | 9.5 |
| Grok 4.5 | 8.9 |
| GPT-5.5 | 8.4 |
| Hy3 | 7.5 |
| Kimi K2.7 Code | 6.5 |
| Composer 2.5 | 4.6 |
| MiniMax M3 | 4.3 |
| DeepSeek-V4-Pro | 4.0 |