Intern-S2-Mobius：解耦知识与推理的基础模型

AI 摘要

Hugging Face 论文页面介绍了 Mobius-v0 架构，该架构将全局共享的 FFN 内存与多个自注意力推理器分离，以提升知识压缩和推理效率。基于此架构，7B 模型仅用基线 62.6% 的训练数据即达到相似性能，而 Intern-S2-Mobius 在持续预训练后实现了近 4 倍的端到端推理加速。

正文节选

Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning Abstract Mobius-v0 separates global memory storage from iterative reasoning modules to improve knowledge compression and inference efficiency, yielding comparable performance with less training data and faster inference. We introduce Mobius-v0, an architecture that comprises a globally shared Memory (FFN) that stores knowledge vectors and multiple Reasoners (Self-Attn) that iteratively achieve compositional reasoning. Usin

发布时间：—

抓取时间：2026-08-17 11:55

来源机构：Hugging Face