Skip to main content

8 Hours to Catch Up, 2 Days to Surpass: How ForgeTrain Uses AI to Build AI Training Frameworks

When large language models evolve faster than most teams can keep up, the real bottleneck isn't just computing power—it's the engineering required to use that power efficiently. On June 27, 2026, Mianbi Intelligence, together with the OpenBMB open-source community and AGI BAR, hosted an event called "AI4AI Fermentation Night." There, Li Yuxuan, technical lead of Mianbi's AI Infra team, unveiled ForgeTrain, a production-grade pre-training framework that flips the script: instead of humans writing training code, AI does it.

The Limits of the Old Way

Li argues that the traditional approach—throwing more data and more GPUs at a problem—has hit diminishing returns. High-quality internet data is getting scarce, and computing power isn't cheap. The industrial revolution gave us machines building machines; the intelligence revolution, he says, is about AI building AI. ForgeTrain is a proof of concept: an automated system that crafts a specialized training framework tailored to a specific model and hardware setup, without any human hand-holding.

Speed That Speaks for Itself

In benchmarks, ForgeTrain matched the performance of Megatron-LM, the industry's gold standard, within 8 hours. Within 1.5 to 2 days, it consistently outperformed it, boosting compute utilization (MFU) by 8% to 10%. And it's not a one-trick pony—the framework has been successfully adapted for models like MiniCPM4-0.5B and 8B, and runs on both NVIDIA H100 GPUs and Huawei Ascend NPUs.

How It Works: The Four-Stage Process

ForgeTrain's magic lies in what Li calls the "Four-Stage Harness Optimization Process." It starts with the Anchor stage, locking down binary consistency. Then comes Bit-for-Bit generation of basic functions. Next, the Surpass stage removes constraints to sprint for performance. Finally, the Per-Op stage dives deep, customizing each operator. The entire pipeline is AI-driven—no human tweaks. This effectively turns NVIDIA's years of engineering know-how into a problem that AI can automatically deconstruct and solve.

Forge Engineering: A New Paradigm

Li dubs this approach "Forge Engineering"—a new way of building software for the AI era. He envisions a future where everyone can customize their own model assistant, and the very form of software will be reshaped. For now, ForgeTrain shows that the path from "AI manufacturing AI" is not just theoretical; it's already delivering results.

Key Points

  • ForgeTrain is an AI-driven framework that automatically generates optimized training code for specific models and hardware.
  • It matches Megatron-LM in 8 hours and surpasses it in 1.5–2 days, with an 8–10% MFU improvement.
  • The four-stage process (Anchor, Bit-for-Bit, Surpass, Per-Op) is fully automated, no human intervention.
  • Compatible with multiple models and hardware platforms, including NVIDIA H100 and Huawei Ascend NPU.
  • Represents a shift toward "AI manufacturing AI," potentially democratizing model training.