KernelBench Mega Normalized (H100)
Code · 2026-09-30
KernelBench Mega H100 PCIe board: best audited Kimi linear-decode cell relative to the current board-best valid cell, scaled to 0–100. Failed/invalid cells contribute zero. The retired RL-grid problem is absent. Best published cells may span harness routes and reasoning settings. Relative to the publisher's H100 comparison pool, not an absolute speedup.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 100.0 |
| Claude Fable 5 | 78.7 |
| Opus 4.8 | 63.8 |
| Kimi K3 | 61.0 |
| GPT-5.5 | 23.1 |
| GPT-5.6 Sol | 18.6 |
| GLM-5.2 | 18.6 |
| Muse Spark 1.3 | 12.4 |
| Gemini 3.5 Flash | 11.3 |
| Grok 4.5 | 10.3 |
| Gemini 3.8 Flash | 8.5 |
| Composer 2.5 | 7.2 |
| DeepSeek-V4-Flash-0731 | 0.0 |
| Kimi K2.7 Code | 0.0 |
| MiniMax M3 | 0.0 |