Physix FrontierPhysix Frontier · News Briefing Card (TMTPost · Nov 5, 2026)
ModelBest agent writes operator in 22.5-hour livestream, 9.8% faster than DeepSeek FlashMLA
KEY FACTS
- ModelBest's ForgeKernel agent wrote an operator 9.8% faster than FlashMLA during a 22.5-hour livestream, with no source code as reference.
- The final version runs in 472 microseconds versus FlashMLA's 518 microseconds, across 42 iterations.
- ForgeKernel topped all three Ascend hardware leaderboards on Huawei's CANNBench, leading the runner-up by 37.1% on the 950DT board.
- ForgeKernel also ranks first in the FlashInfer-Bench track of Nvidia's SOL-ExecBench.
- ModelBest has open-sourced ForgeKernel, with the accompanying paper posted online shortly after.
KEY DATA
9.8%Operator speedup
472 microsecondsFinal runtime
518 microsecondsFlashMLA runtime
37.1%CANNBench 950DT lead
PHYSIX OBSERVATION
An AI autonomously writing an underlying operator that beats a human benchmark means RSI is moving from concept to verifiable engineering practice. For the domestic chip ecosystem, the expert-knowledge moat in software stacks is being chased down by AI on an hourly timescale - a signal more worth watching than model benchmark scores. ModelBest entering the compute substrate at a hundred-person scale also shows that the bar for frontier competition is being redefined.
Source: TMTPost report