China sits at the pivot from manufacturing giant to smart-manufacturing power, and industrial software, CAE simulation, fluid compute and electromagnetic EDA, is the bridge. Domestic industrial software has advanced fast in algorithms and physics modelling toward self-reliance, yet in aerospace, chip design and energy exploration it hits one core bottleneck: not enough speed, it runs slow.
Under the hood of this software, Stencil (stencil compute) is the most basic and most compute-hungry pattern. Because it burns memory bandwidth, unless code meets hardware perfectly, even an advanced GPU stalls waiting for data.
In the past, speeding these up meant a few HPC experts hand-tuning each app over days to weeks, so each expert fixed fewer than 20 apps a year, far behind demand. In the agent age, a Chinese model company is breaking that decades-old lock.
ModelBest with the OpenBMB community open-sourced ForgeStencil, the first AI system to close the loop of automatic research and automatic deploy for Stencil optimisation. In one week it rebuilt and optimised over 100 real industrial and scientific programs, with zero human experts in the loop. It follows ForgeTrain in late May, a fully AI-written pretraining framework 10 per cent faster than Nvidia’s Megatron, and is the latest step in ModelBest’s Forge Engineering paradigm.
Humans only supply the source to optimise. From there ForgeStencil takes over: analyse the app, find the hot function, forge the kernel, swap the operator, verify correctness and reintegrate, with no human in any optimisation decision. It is the first full automatic loop from idea to live deploy.
ForgeStencil proves that rebuilding code through AI can multiply the performance already in the hardware without new spend, and it strikes at the core ask of China’s manufacturing upgrade: replace labour with compute and press fast-forward on domestic industrial software.
Two agents, zero humans
Stencil is among the widest and most basic compute patterns in industrial and scientific software, applying one local operation to each grid point and its fixed neighbours. Atmospheric, seismic, electromagnetic and fluid codes all reduce to it. It is memory bound and often dominates runtime. Get one low-level step wrong and the whole app stalls.
ForgeStencil pairs a Kernel Agent and an App Agent. The Kernel Agent writes and tunes low-level operators toward the hardware limit, generating many candidate kernels across stencil types, grid shapes and precision, trying read patterns, scheduling and structures, then screening the fastest through compile, run and feedback. The App Agent handles integration back into the real app.
Editor’s note: This is an adapted translation of the original OFweek Robot report. It has been trimmed and restructured for readability for an international business audience.