MiMo is a family of large language models from Xiaomi's MiMo team. MiMo-V2.5, published with open weights under the MIT license, is a native omnimodal model: one model that understands text, images, video and audio. It is a sparse mixture-of-experts (MoE) model, a network in which a router sends each token to a few expert sub-networks, so only 15 billion of its 310 billion parameters are active for each token. Its context window holds up to 1 million tokens.
The model mixes sliding-window attention, which looks only at nearby tokens, with global attention; this cuts the size of the KV cache by nearly 6 times. Post-training with reinforcement learning targets agentic tasks, in which the model calls tools over several steps. The weights can be downloaded from Hugging Face and served with SGLang or vLLM. Xiaomi also sells access through the MiMo API on its Open Platform. On 21 October 2026 Xiaomi retired MiMo-V2.5 and MiMo-V2.5-Pro from the API and asked users to move to the V2.6 series.
The API is billed per million input and output tokens, from an account balance; a Token Plan package is also offered.