ModelBest releases fully AI-written pretraining framework ForgeTrain, matching Megatron-LM in 8 hours

面壁智能发布AI全自动预训练框架ForgeTrain，8小时追平Megatron-LM

ModelBest open-sourced ForgeTrain, a pretraining framework whose code is entirely written by AI with zero human edits. It generates model-and-hardware-specific training code from scratch, matches Megatron-LM within 8 hours, and surpasses it after 1.5–2 days with 8–10% higher FLOPS utilization. It has been tested on MiniCPM4-0.5B and 8B, and runs on H100 and Ascend NPUs. The pipeline uses a four-stage automated Harness process. The post does not disclose the open-source license or results on larger models.

Why it matters: ModelBest drops a fully AI-generated pretraining framework that matches Megatron-LM in 8 hours — a hard metric, not a concept demo. Score isn't higher because it's only validated on their own MiniCPM models so far; cross-model generalization data isn't in yet, so I'm discounti...