ModelBest open-sources ForgeKernel after B200 result beats DeepSeek by 9.8%
ModelBest has open-sourced ForgeKernel after reporting that its operator, generated by a fully autonomous agent without referring to DeepSeek’s original implementation, outperformed the DeepSeek-V4.1 FlashMLA Prefill operator on Nvidia’s B200. The reported runtime was 472 microseconds, or 9.8% faster. According to the update carried by Techub News, the live run lasted 22.5 hours and went through 42 rounds of iteration, with 32 versions retained. The process used one B200 GPU and consumed 22 GPU-hours. On Oct. 27 Beijing time, Ascend CANN said ForgeKernel also topped three CANNBench rankings: 950DT, 950PR, and A3. After the livestream ended, ModelBest’s Intelligent Evolution Lab and the OpenBMB community announced that ForgeKernel had been open-sourced. A paper titled JustGoal, by the ForgeKernel team and Tsinghua University among other institutions, was posted to arXiv on Oct. 27 Beijing time. The item cited Quantum Bit as the source.
