ModelBest ForgeTrain: AI-Written Training Framework Matches Megatron-LM in 8 Hours

Forum topic · 小凯 · 2026-07-04

Summary

On July 3, Chinese AI startup ModelBest (Mianbi Intelligence), together with the OpenBMB community and AGI BAR, released ForgeTrain, a production-grade LLM pre-training framework written entirely by AI with no human-written code. According to AI Infra lead Li Yuxuan, ForgeTrain matched the performance of the industry-standard Megatron-LM within 8 hours on MiniCPM4-0.5B/8B models running on NVIDIA H100 and Huawei Ascend NPU hardware, and stably surpassed it within 1.5 to 2 days, achieving roughly 8-10% higher Model FLOPS Utilization (MFU). The framework uses a four-stage 'Harness optimization' pipeline (Anchor, Bit-for-Bit, Surpass, Per-Op) that locks in correctness at each step, progressing from bitwise-identical baselines to deep per-operator tuning. Notably, the same method transfers across NVIDIA and Huawei hardware, suggesting implications for the Chinese domestic NPU software ecosystem. ModelBest frames this as 'Forge Engineering' and positions it at level L2-L3 on its L1-L5 ladder of 'AI building AI.' The post also cautions that the claims cover engineering optimization rather than algorithmic innovation, and that 'no human intervention' means no human wrote code, not that humans were absent from prompting and evaluation.

On July 3, ModelBest (Mianbi Intelligence / 面壁智能), jointly with the OpenBMB community and AGI BAR, released ForgeTrain, a production-grade large model pre-training framework built entirely by AI — with not a single line of code written by humans.

The Headline Numbers

Per technical lead Li Yuxuan, on designated models (MiniCPM4-0.5B/8B) and hardware (NVIDIA H100, Huawei Ascend NPU):

Matched Megatron-LM performance within 8 hours
Stably surpassed it within 1.5–2 days, with MFU (Model FLOPS Utilization) improved by roughly 8%–10%
The same method transfers across NVIDIA H100 clusters and Huawei Ascend NPUs — producing a reproducible engineering approach rather than a vendor-specific implementation

Deep Dive

1. This isn't a new framework — it's the automation of framework-writing

Traditional training frameworks (Megatron-LM, DeepSpeed, PaddlePaddle) rely on humans writing and hand-tuning hardware-specific code. NVIDIA's Megatron-LM team involved dozens of top engineers over five years. ModelBest flipped this: humans write the 'forging rules'; AI writes the optimal framework for the target hardware/model within hours.

2. The core: a four-stage Harness optimization pipeline

Forging is broken into four monotonic, non-rollback stages, each locking the previous stage's correctness:

Anchor: generate a bitwise-identical baseline against the reference implementation (max_abs_diff = 0)
Bit-for-Bit: with binary equivalence guaranteed, quickly implement checkpoint save/load/resume
Surpass: relax binary constraints; let the AI enumerate operator paths, graph capture, and scheduling combos for long-window training stability
Per-Op: deep customization of core ops like GEMM and FlashAttention, maximizing per-op MFU on specific hardware

This turns 'AI writes frameworks' from a black box into a verifiable pipeline — a concrete paradigm within Harness Engineering.

3. 'Forge Engineering' as a new industry term

Li argues that maintaining a general-purpose framework across many hardware backends is increasingly inefficient given the pace of AI model and chip iteration. The alternative: cheap-to-produce, instantly customized training software stacks generated per model/hardware combination. In ModelBest's words, NVIDIA's moat of manpower and ecosystem 'is becoming a technical problem that AI can automatically decouple and reconstruct.'

4. The L1–L5 capability ladder

ModelBest frames 'AI building AI' in five levels: L1 prompt suggestions, L2 assisted R&D, L3 closed-loop delivery, L4 recursive improvement, L5 co-evolution. In general intelligence, initial L4 is reached; but in the vertical 'AI building AI' capability, L2 is solid, and only Anthropic and a very few teams can barely manage unstable L3. ForgeTrain is a single-point empirical push toward crossing L3 toward L4.

Why It Matters

For engineering teams: If this industrializes, the cost structure of LLM training changes — a 5-person team could prompt-generate a training framework optimized for its own hardware instead of running someone else's standard framework. The moat shifts from 'trusted code' to 'trusted intelligence.'

For domestic (Chinese) compute: The forging process demonstrably transfers to Huawei Ascend NPUs. Chinese NPUs have long been criticized for weak software ecosystems; ForgeTrain suggests AI can 'fill in on demand' — a direct engineering tailwind for Huawei, Cambricon, Hygon, Enflame, and others.

For AI safety and governance: Should AI-written code used to train next-generation models fall under compliance review? No standard answer exists yet, but wide adoption of ForgeTrain would force regulators to confront it.

Risks and Caveats

1. Engineering, not algorithmic, innovation. The 8–10% MFU gain and 8-hour parity are optimizations under same-model, same-hardware conditions — the AI did not discover new algorithms. Algorithmic novelty is outside the scope of this L4 single-point experiment. 2. Competitors will accelerate. Megatron-LM's team and big-tech infra groups are likely to be pushed into releasing similar automation tools within months. 3. 'No human intervention' needs careful reading. Experts likely constrained via prompts and validated via experiments; the claim realistically means 'no human wrote code,' not that the process was fully autonomous. Avoid over-hyped framing.

Source: Li Yuxuan (ModelBest AI Infra technical lead) talk at 'AI4AI Fermentation Night,' June 27, 2026, and ModelBest's July 3 recap article.