ForgeStencil: Auto-Optimize 100+ Industrial & Scientific Software Weekly with Zero Human Intervention

Key Highlights

ModelBest (Mianbi AI) together with OpenBMB launched ForgeStencil, claimed as the world's first AI optimization system that supports automatic Stencil research and automatic deployment from a single coordinated workflow. Two agents - a Kernel agent and an App agent - collaborate in a closed loop to fully automate the whole chain from operator optimization to application integration, optimizing over 100 industrial and scientific software packages per week with zero human intervention throughout, which is a striking claim about the scale of automation achieved by the system in real repositories.

Capabilities / What Happened

Stencil computation is one of the most common classes of core operators in scientific computing and industrial simulation (such as fluid, weather, and materials modeling), and has long depended on experts hand-tuning for every new architecture to get acceptable performance. ForgeStencil lets the Kernel agent handle the search and rewriting of low-level operator kernels, while the App agent wires the optimized kernels back into real applications and verifies the effect, forming an optimize, integrate, feedback loop that never stalls waiting for a person. According to reports, the system can already automatically optimize more than 100 industrial and scientific software packages within a week, with no manual intervention required, turning what used to be a multi-person effort spanning months into a background process that runs while engineers sleep. The weekly throughput is only possible because the two agents hand off cleanly: the App agent's verification tells the Kernel agent exactly which kernels still miss the performance target, so the search never wastes time on dead ends that look clever but fail in the real application. This tight feedback loop is what separates ForgeStencil from a batch tuner that optimizes a kernel in isolation and hopes it helps downstream. By keeping the application's actual workload in the loop at every iteration, the system avoids the classic autotuning trap of winning a micro-benchmark while losing the macro one that users actually feel when they run the software. Early adopters note that the biggest surprise is not the speed but the consistency, because the same automated path runs whether the target is a weather model or a fluid solver, removing the heroics that used to accompany each port to new hardware.

Technical Details

The system's key is the closed loop of automatic research plus automatic deployment: the Kernel agent explores high-performance implementations in the search space, while the App agent handles compilation, linking, benchmarking, and regression checks to ensure optimization does not break correctness of the surrounding program. The two agents share evaluation feedback, and errors and performance data flow back into the next search round, continuously approaching better solutions instead of stopping at a local optimum that a single fixed template would lock in. This architecture merges the previously scattered tuning, integration, verification across multiple experts into one repeatable automated pipeline that can be rerun whenever the target hardware changes or a new compiler version appears.

vs. Competitors

Traditional compiler auto-vectorization or hand-written assembly optimization is either general but limited in gain, or efficient but extremely costly, leaving teams to choose between convenience and performance they cannot afford to pursue. ForgeStencil's differentiator is using agents to combine understanding algorithm semantics with connecting application context, rather than blind search alone that ignores how the kernel is actually used by the caller. Compared with generic AutoTuning tools, it emphasizes an end-to-end loop from operator to application, aiming for genuinely deployable, verifiable speedups that survive contact with real workloads rather than benchmark-only wins that evaporate in production code.

Industry Impact / Use Cases

For HPC, industrial simulation, weather, and materials research teams, operator optimization has long been a bottleneck that gates how fast science can move and how big a problem can be attempted on available hardware. ForgeStencil frees up that manpower so scarce experts can focus on higher-level problems instead of micro-optimizing loops by hand late into the night before a deadline. The involvement of OpenBMB, an open-source partner, also means the capabilities have a chance to give back to the community. As such systems mature, AI tuning your core compute for you will move from isolated cases to scaled routine, reshaping who does the tedious but crucial work of making software fast on modern machines. If the claim holds up under independent testing, it suggests a future where performance engineering is a background service rather than a scarce craft, and where porting code to a new chip is mostly a matter of letting the agents rerun their loop overnight. OpenBMB's involvement also raises the odds that the underlying methods get published and reused, rather than kept as a single company's internal advantage that others must reverse-engineer from scratch. If the weekly throughput claim holds, the economics of porting scientific software to new hardware change from a quarterly project into an overnight job. That alone could widen who gets to run cutting-edge simulations, not just those with a dedicated performance engineer already on staff.