ForgeKernel
1 article tagged with ForgeKernel
October 30, 2026
benchmark
ForgeKernel agent beats DeepSeek's FlashMLA Prefill kernel by 9.8% on Nvidia B200 with zero human-written code
ForgeKernel, an operator-generation agent from Mianbi Intelligence (面壁智能), produced a FlashMLA Prefill kernel for DeepSeek-V4.1 that ran in 472 microseconds against FlashMLA's 518 on an Nvidia B200. It did so in a 22.5-hour livestream with no human-written code. The team also open-sourced the system with the OpenBMB community.