ModelBest (面壁智能) and Tsinghua University's NLP lab have jointly open-sourced ForgeTrain, the world's first production-grade large model pre-training framework written entirely by AI. They also released MiniCPM5-1B, a small edge model trained using ForgeTrain. As the first demonstration of a closed-loop "AI builds AI" engineering pipeline, ForgeTrain beats Nvidia's Megatron on the same hardware and delivers a 10% speedup on Huawei's Ascend chips during pre-training. Meanwhile, MiniCPM5-1B has topped the open-weight small model leaderboard on Artificial Analysis.
To let AI autonomously build its own pre-training infrastructure, ModelBest proposed a new software programming paradigm called "Forge Engineering." Instead of a one-size-fits-all framework compatible with every hardware and task, Forge Engineering leverages AI's low-cost code generation to forge specialized code on the fly for specific models and hardware. The construction process involves three stages: first, key data is gathered from existing pre-training frameworks to form a test harness; next, the system iteratively generates binary-compatible framework code in an automated loop; finally, it breaks free from the reference implementation and surpasses it. This whole automated evolution corresponds to levels L3 to L4 in AI building AI.
MiniCPM5-1B, the first model produced by ForgeTrain, has 1.08 billion parameters and is based on a standard LlamaForCausalLM architecture, dramatically lowering the barrier for downstream integration and inference deployment. In Artificial Analysis evaluations, the model scored 18 points, beating the 2B-parameter Qwen3.5-2B (16 points) and outpacing Qwen3.5-0.8B (11 points) and LFM2.5-1.2B-Thinking (8 points). The model supports deployment formats like MLX 4-bit and GGUF Q4_K_M, with INT4 quantization bringing weights down to just 0.5GB. It natively supports a long context of 131,072 tokens and a hybrid dual-mode inference based on enable_thinking. Leveraging its extremely low hardware requirements, OpenBMB also open-sourced MiniCPM Desk Pet, a desktop floating companion app that runs fully offline, responds in real time to coding activity in development tools like Cursor, and supports LoRA persona switching.