KernelBench Hard Normalized (RTX PRO 6000)
Code · 2026-09-30
KernelBench Hard on RTX PRO 6000: 100 times mean cell score divided by the current board-best valid score for each of six problems. Missing, failed and unaudited cells contribute zero. The publisher selects best cells across multiple harness routes/settings for each model, so this is a portfolio aggregate, not a single run. Distinct from the earlier mean raw roofline fraction.
Top models (higher is better)
| Model | Score |
|---|---|
| Kimi K3 | 76.5 |
| DeepSeek V4.1 Flash | 69.6 |
| Claude Fable 5 | 65.5 |
| GLM-5.2 | 63.7 |
| Grok 4.6 | 60.4 |
| GPT-5.6 Sol | 57.8 |
| Grok 4.5 | 47.5 |
| GLM-5.3 | 47.1 |
| DeepSeek-V4-Flash-0731 | 40.4 |
| Opus 4.8 | 37.4 |
| Claude Opus 5 | 33.2 |
| LongCat-2.0 | 21.3 |
| MiniMax M3 | 19.8 |
| Opus 4.7 | 0.0 |
| Sonnet 5 | 0.0 |
| Composer 2.5 | 0.0 |
| DeepSeek-V4-Flash | 0.0 |
| Fugu Ultra v1.0 | 0.0 |
| Gemini 3.5 Flash | 0.0 |
| GLM-5.1 | 0.0 |
| GPT-5.5 | 0.0 |
| Kimi K2.6 | 0.0 |
| Kimi K2.7 Code | 0.0 |
| MiMo-V2.5-Pro | 0.0 |
| Qwen3.6-Max-Preview | 0.0 |
| Qwen3.6 Plus (2026-04-02) | 0.0 |
| Qwen3.7-Max | 0.0 |