Kernel Generation on MUSA KernelBench Level 3
76Pass@8DeepSeek-V4-Pro
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| DeepSeek-V4-ProModel Scale=V4-Pro2026.06 | 76 | 33.8 | 0 | |
| MusaCoder-27B-RLModel Scale=27B, Training Protocol=RL2026.06 | 74 | 54.1 | 3.3 | |
| GLM-5.1Model Scale=5.12026.06 | 70 | 45 | 0 | |
| MusaCoder-9B-RLModel Scale=9B, Training Protocol=RL2026.06 | 54 | 42.5 | 2.6 | |
| MusaCoder-27B-SFTModel Scale=27B, Training Protocol=SFT2026.06 | 46 | 27.9 | 1 | |
| MusaCoder-9B-SFTModel Scale=9B, Training Protocol=SFT2026.06 | 20 | 15 | 0 |