OpenBMB AI Agent Beats NVIDIA FlashMLA Baseline by 9.8% on B200 via Black-Box Optimization

·2026.11.02 23:53

Key point

The agent optimized kernels against a black-box baseline without source code access, releasing the harness and 27 operators.

OpenBMB completed a 22-hour autonomous GPU kernel optimization run on NVIDIA B200, achieving a 1.098x performance improvement (472 µs vs. 518 µs) over the official FlashMLA baseline.
The agent operated in a black-box environment, executing and benchmarking the implementation without access to source code, documentation, or profiling information.
Optimization involved 42 iteration rounds producing 32 kernel versions, with the baseline first surpassed in under 10 hours at version 13.
OpenBMB is open-sourcing the optimized kernel, 27 operators (including attention, Gated DeltaNet, Kimi Delta Attention, GEMM, and convolution), and the ForgeKernel harness for independent experimentation.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.