KernelBench Mega Normalized (RTX PRO 6000)
Code · 2026-09-30
KernelBench Mega RTX PRO 6000 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 cells can combine harness routes/settings; distinct from the earlier absolute speedup metric.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 100.0 |
| Claude Sonnet 5.5 | 90.6 |
| GPT-6 Astra Pro | 70.0 |
| Claude Fable 5 | 69.4 |
| Claude Fable 5.1 | 64.7 |
| GLM-5.3 | 54.8 |
| Kimi K3 | 51.0 |
| DeepSeek V4.1 Flash | 48.2 |
| Opus 4.8 | 40.6 |
| GLM-5.2 | 31.4 |
| Grok 4.7 | 18.0 |
| GPT-5.5 | 12.2 |
| Sonnet 5 | 11.4 |
| Gemini 3.8 Flash | 7.7 |
| GPT-5.6 Sol | 7.4 |
| MiniMax M3 | 7.4 |
| Kimi K2.7 Code | 7.3 |
| GPT-6 Luna | 7.0 |
| Composer 2.5 | 7.0 |
| Gemini 3.5 Flash | 6.4 |
| Muse Spark 1.3 | 6.2 |
| GPT-6 Sol | 2.7 |
| Grok 4.5 | 2.3 |
| DeepSeek-V4-Flash-0731 | 0.0 |
| Grok 4.6 | 0.0 |
| LongCat-2.0 | 0.0 |