Triton kernel generation on KernelBench LEVEL3 1.0
29.8Fast1 ScoreDR. KERNEL-14B-STTS
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DR. KERNEL-14B-STTSscaling_method=sequential test-time scaling (STTS), selection_strategy=best turn selection2026.02 | 29.8 | 7.3 | 0.5 | 0 | |
| GPT-52026.02 | 21 | 12 | 4 | 2 | |
| Claude-4.5-Sonnet2026.02 | 21 | 11 | 5 | 4 | |
| DR. KERNEL-14B-STTSscaling_method=sequential test-time scaling (STTS)2026.02 | 17.1 | 3 | 0.2 | 0 | |
| GPT-5Evaluation Engine=torch.compile2026.02 | 14 | 4 | 3 | 1 | |
| Claude-4.5-SonnetEvaluation Engine=torch.compile2026.02 | 12 | 3.5 | 0.5 | 0 | |
| DR. KERNEL-8B2026.02 | 10.8 | 1 | 0 | 0 | |
| DR. KERNEL-14BEvaluation Engine=torch.compile2026.02 | 9.2 | 3 | 0 | 0 | |
| DR. KERNEL-14B2026.02 | 8.8 | 1.2 | 0.2 | 0 | |
| AutoTriton2026.02 | 7.5 | 0 | 0 | 0 | |
| DR. KERNEL-8BEvaluation Engine=torch.compile2026.02 | 7.2 | 2.3 | 0 | 0 | |
| Qwen3-Coder-A30BA32026.02 | 7 | 1 | 0 | 0 | |
| Qwen3-8B2026.02 | 5.7 | 0.2 | 0 | 0 | |
| GLM-4.72026.02 | 5 | 2 | 2 | 2 | |
| Qwen3-32B2026.02 | 3.5 | 0 | 0 | 0 | |
| Deepseek-V3.2-Thinking2026.02 | 2 | 1 | 0 | 0 | |
| Cold-Start-8Btraining=cold-start data2026.02 | 0.5 | 0 | 0 | 0 |