AI Builds AI: Chinese Company Achieves World First with Self-Written Training Framework
ModelBest (面壁智能) unveils ForgeTrain, the world's first production-grade LLM pretraining framework entirely written by AI, which outperforms NVIDIA's Megatron by 10%. The framework was used to train MiniCPM5-1B, a compact model that sets new records for intelligence density among sub-2B models.
Article intelligence
Key points
- ForgeTrain is the first production-grade LLM pretraining framework fully generated by AI.
- It achieves 10% faster training than NVIDIA Megatron on equivalent hardware.
- The trained model MiniCPM5-1B surpasses all models under 2B parameters on the AA-Index.
- Forge Engineering paradigm shifts from general frameworks to AI-customized code per scenario.
Why it matters
This matters because forgeTrain is the first production-grade LLM pretraining framework fully generated by AI.
Technical impact
May affect model selection, inference cost, product capability, and evaluation benchmarks.
In a landmark achievement for artificial intelligence, Chinese AI company ModelBest (面壁智能) has announced ForgeTrain, the world's first production-grade large language model (LLM) pretraining framework written entirely by AI. This breakthrough not only demonstrates AI's ability to create sophisticated infrastructure but also yields a powerful new model, MiniCPM5-1B, trained using the framework. According to the company, ForgeTrain outperforms NVIDIA's widely used Megatron framework by 10% in training speed on identical hardware, and delivers a 10% speedup on Huawei's Ascend chips compared to existing frameworks.
ForgeTrain was developed using a novel approach called Forge Engineering, which rejects the traditional paradigm of one-size-fits-all frameworks. Instead, AI generates specialized code tailored to specific model architectures, hardware platforms, and training tasks. This is made possible by a three-stage methodology: collecting key data from existing frameworks to build evaluation standards and a harness, constructing a binary-consistent version of the pretraining framework, and then iteratively optimizing beyond the reference implementation. The result is a framework that not only matches Megatron's functionality but exceeds its performance.
The model trained by ForgeTrain, MiniCPM5-1B, is a compact 1-billion-parameter model designed for edge deployment. It can run as a desktop pet or be integrated into mobile devices, laptops, and vehicles. Despite its small size, MiniCPM5-1B achieves high intelligence density: it surpasses all models under 2B parameters on the AA-Index benchmark, and outperforms the recently released Qwen3.5-2B while using half the parameters. The company notes that intelligence density in LLMs is doubling roughly every 3.5 months.
The implications of this development extend beyond technical performance. ModelBest frames AI's ability to build AI as a shift from human-in-the-loop to human-on-the-loop, where researchers design goals and oversight while AI handles the tedious coding. This could accelerate the pace of AI development and reduce costs, especially for Chinese chip ecosystems like Huawei Ascend, where software ecosystem gaps can be narrowed through AI-generated optimizations. ForgeTrain and MiniCPM5-1B have been open-sourced, with links to Hugging Face, GitHub, ModelScope, and other platforms.