Xiaomi MiMo-V2.5 drops 310B MoE

The interesting part is not the headline parameter count; it’s the gap between total size and activated size, which makes this feel much more practical than a brute-force 310B dense model. That said, “more human” is still an inference, not a promise of laptop-class deployment.

• –Sparse MoE plus 15B activated parameters means lower per-token compute, but the full 310B weight stack still keeps deployment serious
• –The 1M-token context and native multimodal stack make it more than a text model with extras; it is aimed at agent workflows, not just chat
• –Xiaomi is clearly building an ecosystem around MiMo, with API, Studio, and adjacent voice models to keep developers inside its stack
• –For local users, the appeal is obvious, but quantization and memory pressure will still decide whether this is workable outside datacenter or high-end workstation setups
• –If the benchmarks hold up in real use, this could become a credible open alternative for multimodal agent work