ModelBest releases ForgeTrain, a fully automated AI pretraining framework that matches Megatron-LM in 8 hours
Original title:AI4AI 技术分享:面壁智能 ForgeTrain 如何 8 小时追平、2 天反超英伟达 Megatron-LM?
AI summary
ModelBest released ForgeTrain, the world's first production-grade large language model pretraining framework written entirely by AI without human intervention. The framework automatically "forges" specialized training code from scratch for specific models and hardware. Benchmarks show that ForgeTrain matches Megatron-LM within 8 hours and achieves a stable lead in 1.5 to 2 days, improving model FLOPS utilization by about 8% to 10%, and can be transferred to different models (MiniCPM4-0.5B/8B) and hardware (H100 and Ascend NPU). It uses a four-stage Harness optimization process with fully automatic decisions throughout. ModelBest summarizes its engineering philosophy as Forge Engineering.
Key Takeaways
- ModelBest released ForgeTrain,the world's first production-grade large language model pretraining…
- The framework automatically "forges" specialized training code from scratch for specific models…
- Benchmarks show that ForgeTrain matches Megatron-LM within 8 hours and achieves a stable lead in 1.
- 5 to 2 days,improving model FLOPS utilization by about 8% to 10%,and can be transferred to…
- 5B/8B) and hardware (H100 and Ascend NPU).
Quick Facts
- Published
- Updated
- Source
- 公众号:面壁智能(MiniCPM)
- Category
- Products & Apps
- AI score
- 65
这是AI编写生产级训练框架的首个实证,8小时追平Megatron-LM并反超,证明AI打穿Infra的范式已到临界点,做训练框架和Infra的都该看看。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: ModelBest releases ForgeTrain, a fully automated AI pretraining framework that matches Megatron-LM in 8 hours. July 3, 2026. https://new.nat6.net/en/news/cmr4righa04d8sll54465w8z3/