Benchmarks · coding
KernelBench Hard (kernelbench.com)
Coding agents write six GPU kernels (FP8 and W4A16 GEMMs, KDA, paged attention, top-k, MoE) in CUDA or Triton, one unlimited session per problem on an RTX PRO 6000, each run audited for reward hacking. Mean peak fraction of roofline over the full deck; failed, flagged or unaudited problems count 0. Independent of Stanford's KernelBench.
11 models · 11 results · updated 2026-10-18 · kernelbench.com ↗
Leaderboard
Grok 4.6 xhigh
More from this benchmark
9 published measuresScore vs any published measure9 measures: sub-scores, splits, counts
Breakdown
darker is better within each column · click a column to sort| Model | ||||||
|---|---|---|---|---|---|---|
| Grok 4.6 xhigh | 0.375 | 0.0406 | 0.622 | 0.0368 | 0.234 | |
| DeepSeek V4.1 Flash 1017 max | 0.324 | 0.0403 | 0.459 | 0.0897 | 0.251 | 0.106 |
| Fable 5 max | 0.391 | 0.43 | 0.0492 | 0.261 | 0.105 | |
| Qwen 3.8 Max xhigh | 0.327 | 0.693 | 0.0397 | 0.107 | ||
| DeepSeek V4.1 Flash max | 0.331 | 0.0374 | 0.362 | 0.0425 | 0.249 | 0.108 |
| GLM 5.3 max | 0.632 | 0.0885 | 0.365 | |||
| DeepSeek V4 Flash 0731 max | 0.409 | 0.486 | 0.0285 | 0.156 | ||
| Kimi K3 max | 0.32 | 0.0315 | 0.485 | 0.064 | 0.0885 | |
| GPT-5.6 Sol xhigh | 0.387 | 0.0412 | 0.198 | |||
| Opus 5 max | 0.0946 | 0.371 | ||||
| GLM 5.3 Flash max | 0.0717 | 0.102 |
All results
11 rows| # | Model | Effort | Score | Harness |
|---|---|---|---|---|
| 1 | Grok 4.6 | xhigh | 21.8% | grok |
| 2 | DeepSeek V4.1 Flash 1017 | max | 21.2% | or-fable |
| 3 | Fable 5 | max | 20.6% | or-fable |
| 4 | Qwen 3.8 Max | xhigh | 19.4% | or-fable |
| 5 | DeepSeek V4.1 Flash | max | 18.8% | deepseek-claude |
| 6 | GLM 5.3 | max | 18.1% | zai-claude |
| 7 | DeepSeek V4 Flash 0731 | max | 18.0% | or-fable |
| 8 | Kimi K3 | max | 16.5% | kinetic-claude |
| 9 | GPT-5.6 Sol | xhigh | 10.4% | codex |
| 10 | Opus 5 | max | 7.8% | or-opus |
| 11 | GLM 5.3 Flash | max | 2.9% | or-fable |