When the iteration speed of large models is measured in weeks, how to break through the computing power bottleneck and engineering barriers has become the ultimate question in AI development. On June 27, 2026, Mingshi Intelligence, in collaboration with the OpenBMB open-source community and AGI BAR, held an in-depth exchange event named "AI4AI Fermentation Night." During the event, Li Yuxuan, the technical lead of Mingshi Intelligence's AI Infra, provided a detailed analysis of its self-developed production-grade pre-training framework - ForgeTrain, revealing the underlying logic and practical breakthroughs of the paradigm shift "AI manufacturing AI."
Li Yuxuan pointed out that as the marginal benefits of high-quality internet data and computing power supply are diminishing, the traditional path of "stacking data and computing power" has reached its limit. The industrial revolution realized "machines manufacturing machines," while the intelligent revolution is moving toward "AI manufacturing AI." ForgeTrain is a single-point verification of this path. Its core idea is: using AI to forge a specialized training framework tailored for specific models and hardware architectures from scratch, rather than relying on manually maintained general software stacks.
In performance tests, ForgeTrain demonstrated remarkable iteration efficiency. Through an automated process, the framework matched the performance of the industry flagship framework Megatron-LM within 8 hours and achieved stable surpassing within 1.5 to 2 days, with an MFU (compute utilization) increase of about 8% to 10%. This capability has successfully been migrated to different models such as MiniCPM4-0.5B/8B and is compatible with various hardware platforms, including H100 and Huawei Ascend NPU.
The success of ForgeTrain relies on Li Yuxuan's summarized "Four-Stage Harness Optimization Process." Starting from the Anchor stage, where binary consistency is locked, it goes through Bit-for-Bit basic function generation, Surpass performance sprint after removing constraints, and finally enters the Per-Op stage, which involves deep customization of each operator. The entire process is entirely determined by AI without any manual intervention, successfully transforming NVIDIA's years of accumulated engineering moat into a technical problem that can be automatically decoupled by AI.
This practice is summarized as "Forge Engineering" - a new engineering paradigm for the AI era. Li Yuxuan believes that in the future, everyone will have the ability to customize their own model assistant, and the software form will undergo a large-scale reshaping.
