@testingcatalog· @testingcatalog · X·· 2026-09-03精选AI 评分70
AI 导读
IFM 在发布模型的同时公开了训练代码、数据配方、中间 checkpoint、日志与评测结果。随模型一同发布的两个架构组件是：MoVA 在注意力内部路由专家，Uno 是可并行生成 token 块的扩散适配器。权重与代码已通过链接给出。
推荐理由
原文给出模型权重之外还公开训练代码与两个架构组件，读者可据此判断这次开源发布的具体范围。
正文
Together with the models, IFM also published the training code, data recipes, intermediate checkpoints, logs, and evaluations behind them.
Two architecture pieces ship with the models:
> MoVA routes experts inside attention itself
> Uno is a diffusion adapter that can generate token blocks in parallel.
Weights and code 👇
https://t.co/4KR5rbPmoq
来源：@testingcatalog · x.com