The world's first AI pre training framework is open source, and Tsinghua University and Face Wall have launched ForgeTrain According to Beating monitoring, Face Wall Intelligence and Tsinghua NLP Laboratory jointly open sourced the world's first fully AI written production grade large model pre training framework, ForgeTrain, in the OpenBMB community, and released the end side small model MiniCPM5-1B trained by ForgeTrain. As the first sample to showcase the closed-loop of "AI Manufacturing AI" engineering, ForgeTrain outperforms Nvidia's Megatron in performance under the same hardware conditions, and also achieves a 10% acceleration during pre training on Huawei Ascend. At the same time, MiniCPM5-1B topped the Artificial Analysis Open Weight Small Model List. In order to enable AI to independently build underlying pre training infrastructure, Face Wall Intelligence proposes the "Forge Engineering" software programming paradigm, abandoning a universal framework that is compatible with all hardware and tasks, and instead utilizing AI's low-cost code generation capabilities to forge specialized code for specific models and hardware on-site. In terms of construction mechanism, ForgeTrain adopts a three-stage approach: first, key data is collected from existing pre training frameworks to form a test room (Harness), then binary consistent framework code is iteratively generated in an automatic closed loop, and finally the restrictions are lifted and the reference implementation is surpassed. The entire automation evolution corresponds to the L3 to L4 stages of AI manufacturing. As the first output model of ForgeTrain, MiniCPM5-1B has 1.08 billion parameters and its core architecture is based on the standard LlamaForCauselLM design, significantly reducing the threshold for downstream integration and inference deployment. In the Artificial Analysis evaluation, the model surpassed the 2B scale Qwen 3.5-2B (16 points) with a score of 18, and led Qwen 3.5-0.8B (11 points) and LFM 2.5-1.2B-Thinking (8 points). The model supports deployment formats such as MLX 4-bit and GGUF Q4_K_M, with INT4 quantized weights of only 0.5GB. It also natively supports 131072 token long text context and hybrid dual-mode inference based on enable_thinking. With extremely low hardware costs, OpenBMB has simultaneously opened sourced the desktop floating window companion application MiniCPM Desk Pet, which runs purely offline. It supports real-time response to coding activities and LoRA persona switching in development tools such as Cursor. [Original link]