Shanghai AI Lab Presenta Intern-S2-Mobius: MoE de 397B con Atención Desacoplada de Memoria

The Shanghai AI Laboratory has published the Intern-S2-Mobius family, spearheaded by the 397B MoE preview model and the innovative Intern-MemDec-4B memory-decoupling system.

Memory-Decoupled Attention (MemDec)

Intern-S2 separates working memory from historical associative retrieval. By utilizing a dedicated 4B auxiliary network to compress conversational context into compact memory vectors, the primary 397B MoE focuses solely on active mathematical deduction and theorem proving.

In tandem with InternLumina-U2 multimodal diffusion checkpoints, Intern-S2 sets high-water marks on OlympiadBench mathematics and formal Isabelle theorem proof generation.