Xiaomi MiMo-V2.5

MiMo-V2.5 is built for long, tool-driven work that may mix software code with visual, video, and audio information. Its sparse design activates 15 billion of 310 billion parameters per token, seeking a balance between broad model capacity and practical inference efficiency. A three-layer multi-token prediction module can accelerate decoding, while EAGLE speculative decoding can use the accompanying MTP weights.

The architecture interleaves sliding-window and global attention to limit key-value cache growth while retaining access across long inputs. This foundation suits complex coding, multistep reasoning, document analysis, and agent workflows that must interpret multiple media types. Its native multimodal path uses a vision encoder and audio transformer, and the documented deployment provides multimodal understanding through an OpenAI-compatible interface.