ForgeKernel agent beats DeepSeek's FlashMLA Prefill kernel by 9.8% on Nvidia B200 with zero human-written code

An autonomous agent called ForgeKernel wrote an attention kernel that ran 9.8% faster than DeepSeek's hand-written FlashMLA Prefill operator on an Nvidia B200, according to a livestreamed run reported by QbitAI (量子位), a Chinese-language outlet. The final kernel took 472 microseconds against 518 microseconds for FlashMLA. The agent received only the operator's mathematical definition and target hardware constraints, with no source code and no reference to DeepSeek's implementation.

ForgeKernel comes from Mianbi Intelligence (面壁智能). Its Intelligent Evolution Lab announced it would open-source the system together with the OpenBMB community after the livestream ended.

The livestream run

The target was the FlashMLA Prefill operator for DeepSeek-V4.1. FlashMLA is DeepSeek's open-source attention kernel library, and DeepInfer, the inference framework DeepSeek open-sourced on October 20, integrates it. The run used one B200 and 22 GPU hours across a 22.5-hour broadcast.

• First version: passed correctness checks at 1834 µs.
• Under 4 hours: 593 µs, a drop of about 67.6%.
• Around hour 7: a core structural change let it beat its own previous version, then FlashMLA for the first time.
• Final: 472 µs versus FlashMLA's 518 µs, across 42 iterations and 32 saved versions.

Late in the run, the agent found that bursty Q loads competed with KV data movement for memory bandwidth. It throttled the Q request rate, cutting a further 2.4% of runtime.

Ascend leaderboard results

On October 27, Huawei's Ascend CANN account said ForgeKernel topped three CANNBench leaderboards:

Hardware	Score	Speedup vs baseline	Versus second place
950DT	4010.75	1.533x	27.8% ahead on score
950PR	4165.94	1.638x	about 1.18x
A3 (Ascend 910C)	4341.16	1.898x	about 1.22x

The A3 result has held since CANNBench v1.1 closed on October 15. The same system ran on CUDA for the B200 and on Ascend's separate toolchain. According to the team, moving between them required swapping the profiler from NCU to msprof.

Method and broader results

On October 27 the team and Tsinghua University and other institutions posted a paper on arXiv titled "JustGoal: With Only a Goal, the AI Baseline Is Above the Expert Kernels." Given only an optimization goal, the model beat human-maintained open-source libraries on 8 of 10 operators on B200. The same model inside the KDA-0.5 framework, which MIT's Han Song team helped develop, scored 0.95x of JustGoal on average for single-configuration tasks.

The paper found that models tune details well but struggle to design overall architecture, often locking in a route within the first few attempts. ForgeKernel addresses this in three parts:

1. It separates architecture design from implementation and expresses design as milestones.
2. It stores each hardware lesson validated by measurement in a knowledge base for reuse on other operators.
3. It treats its own best version as the opponent for the next round.

Across 29 comparisons, ForgeKernel matched or beat expert implementations in 24. It also matched or beat all 7 strongest operators published by three leading agent systems.

Deployment

ForgeKernel-generated operators are integrated into a training framework and completed a long training test of MiniCPM5-1B on domestic Chinese chips and Nvidia platforms. The team says applying the approach to a training framework gave 20% to 30% speedups on domestic chips and 5% to 10% on Nvidia. DeepSeek V4.1 is already being used to help generate and optimize operators, per QbitAI.

Mianbi's chief scientist for intelligent evolution, Li Yuxuan, said that rewriting top kernels such as the FlashAttention series used to take human experts months to half a year per iteration. Internally, he said, an AI task that surpasses the best human implementation takes at most two days.

What this means

The livestream shows a measured result on one operator and one GPU: an agent starting from a math definition produced a Prefill kernel faster than FlashMLA's 518 µs. The broader claims rest on the team's own paper and the CANNBench leaderboards. The reported 24 of 29 matched or better comparisons indicate the result is not confined to a single kernel. The profiler swap between B200 and Ascend suggests the optimization process is not tied to one chip, though kernel implementations themselves remain hardware-specific. Open-sourcing lets outside teams test the system against their own operators.