Zero Human Code, 22.5 Hours: An AI Agent Wrote a GPU Kernel 9.8% Faster Than DeepSeek's FlashMLA

A Chinese research team has published the closest thing yet to a clean answer to an uncomfortable question: if a general-purpose agent is handed nothing but a performance target, can it out-code the best humans in GPU kernel engineering? On October 30, ModelBest put that question on a public livestream and let it run for 22.5 hours. The agent, called ForgeKernel, was given the mathematical definition of DeepSeek's FlashMLA Prefill attention operator and the constraints of an Nvidia B200 GPU. It was not given DeepSeek's source code, and no human wrote a single line for it.

The final artifact ran in 472 microseconds. DeepSeek's FlashMLA implementation, widely treated as the high-water mark of hand-tuned attention kernels, ran in 518 microseconds on the same hardware. That is roughly 9.8 percent faster, produced by 42 iterations that left 32 saved versions behind, burning 22 GPU-hours on a single B200.

The trajectory matters more than the endpoint. The first version passed correctness checks at 1,834 microseconds, well off the pace. Within four hours the agent had cut that to 593 microseconds, a 67.6 percent reduction that already put it at about 90 percent of the best existing implementation. Around hour seven a structural rewrite pushed it past its own previous best and then, for the first time, past FlashMLA. In the closing hours it noticed on its own that burst loads of the Q tensor were competing with KV data movement for memory bandwidth, throttled the Q request pacing, and clawed back another 2.4 percent. That is the kind of call that normally requires a senior performance engineer with years of hardware intuition.

ForgeKernel comes out of ModelBest's Intelligent Evolution Lab and the OpenBMB community, and its design splits the problem in three. Architecture decisions are separated from implementation and broken into milestones, so the agent does not lock itself into a dead-end structure in its first few attempts. Every hardware lesson that survives measurement goes into a knowledge base that carries over to the next operator. And each round competes against the agent's own best previous version, which is what produced the self-improvement curve visible in the livestream.

The method is documented in a paper posted to arXiv on October 27 by researchers from Tsinghua University, ModelBest, Harbin Institute of Technology and the University of Chinese Academy of Sciences. Titled "JustGoal: With Only a Goal, the AI Baseline Is Above the Expert Kernels," it deliberately strips out everything the specialized kernel agents usually rely on: the evaluation harness stays, but domain knowledge and prescribed search workflows are removed. On B200 the goal-only baseline beat human-maintained open-source libraries on 8 of 10 operators, with per-operator geometric mean speedups of 1.14x to 2.34x, and beat the fastest publicly released specialized-agent kernel on 5 of the 9 operators where a comparison exists. In a controlled same-model test, MIT's KDA-0.5 harness reached only 0.95x of the goal-only result. The repository is on GitHub under Apache 2.0.

The result also travels across architectures. On October 27, Huawei's CANN team reported that ForgeKernel took first place on all three CANNBench boards for Ascend silicon: 950DT with 4,010.75 points, 27.8 percent ahead of second place at a 1.533x baseline speedup; 950PR with 4,165.94 points and 1.638x; and the widely deployed A3, or Ascend 910C, with 4,341.16 points and 1.898x. The agent does not carry a preset workflow for any specific chip. Moving from Nvidia to Ascend meant swapping the profiling tool from NCU to msprof and letting the model re-learn the hardware by measuring it.

The economics are the point. Gartner puts 2026 AI infrastructure spending at roughly 1.48 trillion dollars, so a one percent gain in compute efficiency is worth about 14.8 billion dollars at that scale. The team says the same approach yields 20 to 30 percent speedups when optimizing training frameworks on domestic Chinese chips, against 5 to 10 percent on Nvidia hardware, where years of CUDA work have already squeezed out most of the easy wins. In 29 head-to-head comparisons ForgeKernel matched or beat expert implementations in 24, and matched or beat all seven of the strongest publicly released kernels from three leading agent systems.

None of this is only a benchmark story. The generated kernels have been wired into a training framework and used for long-run training of MiniCPM5-1B on both domestic accelerators and Nvidia platforms, and the whole system was open-sourced after the livestream ended. Li Yuxuan, chief scientist at ModelBest's Intelligent Evolution Lab, frames the shift bluntly: top-tier kernels such as the FlashAttention family used to take human experts months, sometimes half a year, per iteration, while inside the team an agent now beats the best human implementation in at most two days. The caveat is that every figure here is team-reported, and the livestream number has not been independently reproduced; the paper's code is open precisely so that others can try. But the direction is hard to argue with. When writing a kernel costs almost nothing, optimizations that were never worth an engineer's salary start to pencil out, and the moat around a mature software ecosystem stops being measured in expert-years.