MiMo-V2.5

MiMo-V2.5 is Xiaomi's native omnimodal mixture-of-experts model, designed as a single architecture that can absorb text, images, video, and audio without bolting on separate pipelines. It carries 310B total parameters with 15B active per token, sitting on a 48-layer backbone made of one dense layer and 47 MoE layers. The routing uses 256 experts with a top-8 selection, paired with a hybrid attention scheme that alternates sliding-window attention over a 128-token window with full attention at a 5:1 ratio. Dedicated encoders handle the non-text streams: a 729M vision encoder and a 261M audio encoder feed into the same backbone, and a three-layer Multi-Token Prediction head sits on top of the stack. The model ships in native FP8 block-wise e4m3 weights, which keeps the footprint manageable enough that community reports note it can run across a pair of high-end desktop accelerators. Open weights make the whole package available for self-hosting and fine-tuning rather than locking users into a single provider's stack.

MiMo-V2.5 inherits its backbone from the MiMo-V2-Flash line, and the broader V2.5 generation was tuned to push agentic capability and long-horizon coherence. On Xiaomi's published Coding Agent benchmark, MiMo-V2.5 lands at 56.1, and it reaches 71.8 on SWE-Bench Pro, putting it ahead of the previous V2-Pro generation and within striking distance of much larger frontier systems. The release notes emphasize complex software engineering and multi-step workflows as the primary intended use cases, and the model is wired for tool calling and structured reasoning out of the box. Hosting providers benchmark throughput north of 130 tokens per second, which positions MiMo-V2.5 as a practical option for production agents that need both multimodal grounding and fast iteration. For builders looking ahead, the combination of open weights, a compact active-parameter count, and native multimodal encoders makes it a flexible base for everything from autonomous coding assistants to video-and-audio-aware enterprise tools.