ModelBest's ForgeKernel Beats DeepSeek's FlashMLA by 9.8% in a 22.5-Hour Live Run, Then Open-Sources

A 22.5-hour livestream ended on the morning of October 30 with an agent that had no human involvement. Given only a kernel's mathematical definition and target hardware constraints, and no DeepSeek source code or reference implementation, it rewrote the kernel through repeated self-improvement until it ran faster than DeepSeek's latest FlashMLA Prefill kernel.

The numbers were visible throughout. The first version passed correctness checks but took 1,834 microseconds; within four hours it was down to 593 microseconds, a 67.6% cut and roughly ninety percent of the best existing implementation; around hour seven, a structural breakthrough pushed it past FlashMLA for the first time; at the end, the final version ran in 472 microseconds against 518 for FlashMLA. The run used 42 iterations, produced 32 versions, consumed one B200 and 22 GPU hours, and involved zero lines of human-written code.

The target was not a soft one. FlashMLA is DeepSeek's open-source core attention kernel library, integrated into DeepInfer, the inference framework DeepSeek released on October 20. It comes from a team widely regarded as among the strongest in infrastructure and has stood as a high-water mark for hand-written kernels.

The livestream was not ForgeKernel's first appearance. On October 27, Huawei's Ascend CANN team published a post naming ForgeKernel as the top entry on all three CANNBench hardware leaderboards: 950DT, 950PR and A3. It scored 4,010.75 on 950DT, 27.8% ahead of second place with a 1.533x baseline speedup; 4,165.94 on 950PR with 1.638x; and 4,341.16 on the A3 board for the widely deployed Ascend 910C, with 1.898x.

950DT is a newer chip whose companion kernel library is still early, and manual adaptation is slow. ForgeKernel adapted first and topped the board, which suggests the gap between a chip shipping and its kernels being ready is narrowing. More tellingly, the livestream ran on an NVIDIA B200 with CUDA while the leaderboards ran on Ascend's own architecture and toolchain, and one system took first place across both.

On method, the ForgeKernel team and institutions including Tsinghua University posted a paper called JustGoal to arXiv on October 27. The setup gives the model only an optimization goal, with no kernel knowledge and no fixed workflow. According to the paper, the model beat human-maintained open-source libraries on eight of ten kernels on a B200.

After the livestream, the Intelligent Evolution Lab and the OpenBMB community announced that ForgeKernel would be open-sourced. That turns a demonstration into a reproducible engineering artifact: outside teams can test the self-evolution loop against their own chips, models and kernel definitions instead of watching a replay.

Kernel tuning has long depended on senior engineers working one kernel at a time, and chip and model release cycles now outpace the supply of that expertise. Read through the leaderboard results, the same hardware spend yields nearly 50% more usable compute, with procurement and deployment costs potentially halved. The test comes after open-sourcing: whether the community can reproduce the results, whether generated kernels hold up under real training and inference loads, and whether the approach changes how quickly domestic chip software stacks are adapted.