Xiaomi

Xiaomi Corporation, a consumer electronics and smart-manufacturing company founded on April 6, 2010, whose products include smartphones, electric vehicles and smart hardware,¹ also builds language models under the MiMo name. They come from its LLM-Core team.⁴ The first, MiMo-7B, went up on Hugging Face on April 29, 2025 as a series of 7-billion-parameter base, fine-tuned and reinforcement-learning models “trained from scratch and born for reasoning tasks”, pre-trained on about 25 trillion tokens Vendor-stated, with a technical report posted to arXiv on May 12, 2025.²³⁴ Vision-language, audio and embodied MiMo models followed later in 2025.²

The series moved to a mixture-of-experts (MoE) design, in which each token passes through only part of the network, with MiMo-V2-Flash, a 309-billion-parameter model released in December 2025. Using OpenRouter data on the top 10 open models, the ATOM Report finds it went from zero to 27.2% of open-model inference tokens by January 2026, behind only DeepSeek at 31.1% Measured.⁵ The current generation, MiMo-V2.6, was published on Sep 21, 2026. Its card describes MiMo-V2.6-Flash as natively omnimodal across text, image, video and audio, with 309 billion total and 15 billion activated parameters and a 1M-token context, and the weights are MIT-licensed Vendor-stated.⁶²

On OpenRouter, an API service that routes requests to hundreds of models, MiMo-V2.6-Flash processed 11.54 trillion tokens in the seven days to Oct 12, 2026, fourth among all models and up 19% on the week before Measured, as ranked in Most-Used AI Models on OpenRouter by Weekly Tokens.⁷ On Hugging Face, Xiaomi’s most-downloaded general language model on Oct 13, 2026 was MiMo-V2.5, with 192,784 downloads in the previous 30 days Measured, one row in Chinese Open-Weight Models: Downloads, Derivatives and Token Share.²