China is crossing from “manufacturing giant” to “smart-manufacturing power,” and industrial software, CAE simulation, fluid dynamics and electromagnetic EDA, is the core bridge. Domestic tools have advanced fast in algorithms and physics modelling, but at the front line of aerospace, chip design and energy, they hit one bottleneck: not smart enough, just slow.
Under the hood of this software, Stencil is the most basic and most compute-hungry pattern. Hungry for memory bandwidth, it stalls even advanced GPUs if code and hardware are not perfectly matched. Historically, speeding it up meant a few HPC experts hand-tuning for days to weeks, optimising under 20 apps a year, far behind demand.
In the Agent age, that decades-old lock is breaking. ModelBest, with the OpenBMB community, open-sourced ForgeStencil, the first AI system with a full “auto-research plus auto-deploy” loop for Stencil optimisation. In one week it refactored and optimised 100-plus real industrial and scientific apps, with zero human experts in the loop.
It follows ForgeTrain, released in late May as the first fully AI-written pretraining framework, 10 per cent faster than Nvidia’s Megatron. ForgeStencil is the next step in ModelBest’s “Forge Engineering” paradigm.
Two agents, zero humans
Humans only supply the source to optimise. ForgeStencil takes over: analyse the app, locate hot functions, forge kernels, swap operators, verify correctness and integrate back, with no expert in any decision. That is the first full loop from “optimisation idea” to “seamless deploy.”
ForgeStencil proves AI can rebuild code and squeeze existing hardware without new investment, hitting China’s core upgrade need: replace labour with compute, and press the self-evolution button on domestic industrial software.
The system has two agents. Kernel Agent writes and optimises low-level operators toward the hardware limit, generating candidates, trying read patterns, thread schedules and compute structures, then screening the fastest by compile-run-feedback loops. App Agent handles the application side.
Stencil matters because it is everywhere: atmospheric dynamics, seismic waves, electromagnetic fields, fluid and phase-field simulation all reduce to updating a grid by weighted neighbours. Memory-bound, it dominates runtime. ForgeStencil breaks the expert bottleneck that could never scale.
Editor’s note: This is an adapted translation of the original OFweek Robotics report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://robot.ofweek.com/2026-08/ART-8321205-8120-30697117.html.