Kernel Arena - WaferBench NVFP4: leaderboard

Kernel Arena: Cursor-harness runs on six NVFP4 fused-inference kernel tasks for NVIDIA B200. The score is mean correctness plus speedup versus FlashInfer references; incorrect kernels score zero. Scores come from Wafer’s official suite exports.

Metric: Mean Correctness+Speedup. Source: github.com. Saturation forecast: Around July 2027. 4 models tracked.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)2.27#58
2GPT-5.42.18#79
3Claude Opus 4.62.03#63

No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.

Interactive version: theaggregate.ai/benchmark?slug=kernel-arena-waferbench-nvfp4 · How It Works · Data refreshed daily, snapshot 2026-10-20.