मिमो v2.5

Xiaomi's natively omnimodal model, handling text, image, video and audio in one unified architecture. It is a sparse Mixture-of-Experts with 310B total and 15B active parameters, pairing a vision encoder and an audio transformer with the language backbone, and it inherits a hybrid sliding-window attention design that cuts KV-cache storage nearly sixfold. It supports up to a million tokens of context at a very low price.

Xiaomi built this one to perceive and to act in the same pass. The language backbone is inherited from the team's earlier hybrid sliding-window model, and the vision and audio encoders - both pretrained in-house - hang off it through lightweight projectors rather than being bolted on as a separate pipeline. The lab's summary of the goal is a single model that sees, hears and acts on what it perceives, and the training schedule is built around that: text pre-training for the backbone, a projector warmup to align audio and vision with the language model, large-scale multimodal pre-training, supervised fine-tuning and agentic post-training during which the window is stretched from 32K to 256K to a million tokens, and finally reinforcement learning with multi-teacher on-policy distillation.

The agent claims are where Xiaomi puts its emphasis, and it argues them on efficiency as much as accuracy. On its internal coding evaluation it reports the model matching its own far larger Pro sibling at half the cost. On its daily-agent benchmark it reports 62.3 on the general subset and places the result at the frontier of score against token spend, which is the whole pitch: frontier-level agent behaviour without frontier-level token bills. On the public coding sets it reports 65.8 on Terminal-Bench 2.0 and 56.1 on SWE-Bench Pro.

The perception numbers are the other half. Xiaomi reports 81.0 on the chart-reasoning benchmark CharXiv, 77.9 on MMMU-Pro, 88.5 on high-resolution image understanding, 87.2 on document understanding, 87.7 on general video question answering and 64.0 on the harder video-reasoning set VideoHolmes. Its framing of those is comparative and specific: level with a leading closed model on video, level with another on multimodal agentic work, and competitive rather than leading on image and document understanding.

Where it stops shows up in the same table. On the multimodal agent split - real, messy interactions rather than single questions - Xiaomi reports 23.8, behind two of the closed models it compared against, and the number is low in absolute terms for every model on that row. The lab also positions the larger Pro model above this one for the hardest long-horizon software engineering. Read this as the omnimodal workhorse: cheap, very long-context, strong at everyday agent work and at reading what it is shown, with the hardest autonomous runs left to its bigger sibling.