Model Releases
I have some very big news... KernelBench-Hard with H100 and B200 (single gpu results) AND KernelBench-Mega tested on RTX PRO 6000, H100, B20…
I have some very big news... KernelBench-Hard with H100 and B200 (single gpu results) AND KernelBench-Mega tested on RTX PRO 6000, H100, B200 is finally out! Starting with Mega, each of models wrote a
I have some very big news... KernelBench-Hard with H100 and B200 (single gpu results) AND KernelBench-Mega tested on RTX PRO 6000, H100, B200 is finally out! Starting with Mega, each of models wrote a GPU megakernel (that means one CUDA kernel per token generated) from scratch, on three NVIDIA GPUs (RTX PRO 6000, H100, B200), and open-sourced every agent trace. Claude Opus 4.8 wins on every GPU, up to 19.4x over the reference on B200. GLM-5.2 is the top open-weight model and its not close! Full results + 172 traces below if you want to review/train on them. HUGE thank you to @NVIDIAAI for sponsoring me credits to run this on datacenter hardware!
Source: Clem Delangue (X) | 2026-06-20