Xiaomi, the Beijing company best known for smartphones, software and electric vehicles, released its first MiMo language model, the MiMo-7B reasoning model, in April 2025. Most MiMo weights since then are open under the MIT licence, up to a 1.02 trillion parameter model, but some MiMo releases have been API-only.
What MiMo is
Wikipedia describes MiMo as Xiaomi’s family of LLMs and the central AI model in the company’s “Human x Car x Home” ecosystem, with developers reaching it through an API. Xiaomi itself was founded in 2010 by Lei Jun with six associates. Wikipedia reports that Lei Jun said Xiaomi planned to invest at least US$8.7 billion in AI over three years.
Release history
MiMo-7B, from April 2025, is a reasoning model, and its card lists base, SFT and RL checkpoints. Wikipedia lists two siblings: MiMo-VL-7B, a vision-language model, and MiMo-Audio-7B, an audio-language model.
MiMo-V2-Flash came in December 2025: a mixture-of-experts model that its card says is designed for high-speed reasoning and agentic workflows.
MiMo-V2-Pro was not open at launch. Wikipedia reports that Xiaomi introduced it on 18 March 2026 as a proprietary API product, after it had appeared anonymously on OpenRouter as “Hunter Alpha”. MiMo-V2-Omni and MiMo-V2-TTS came the same day, also proprietary. Wikipedia quotes Luo Fuli saying Xiaomi intended to open-source a variant of V2-Pro at an unspecified date.
The open release followed in April 2026 with MiMo-V2.5 and V2.5-Pro. The V2.5-Pro card calls it an open-source mixture-of-experts model with up to 1M tokens of context, aimed at demanding agentic and software engineering tasks.
Wikipedia lists MiMo-V2.6 in Pro and Flash versions, released on 21 September 2026.
MiMo-V2.6-Distill-Qwen-9B is the odd one out. Its card names Qwen/Qwen3.5-9B as its base, with a “finetune” relation, and describes distillation and supervised fine-tuning. Qwen3.5-9B lists Apache 2.0 on its own card, so check both.
Licences
The cards for MiMo-7B-RL, MiMo-V2-Flash, MiMo-V2.5-Pro, MiMo-V2.6-Pro-RL and the Qwen-based 9B model all list the MIT licence, and we found no switch to a restricted licence between releases. The caveat is the API-only models: Wikipedia says MiMo-V2-Pro, V2-Omni and V2-TTS were proprietary, with no public weights for TTS. The MIT licence page covers what MIT allows.
Where it fits among Chinese labs
Xiaomi is a hardware company that also trains and releases large language models. Its peers have different backers: DeepSeek is funded by a hedge fund, while Moonshot AI, Zhipu and MiniMax are AI start-ups with outside investors (see who owns them). Wikipedia names Xiaomi itself among the investors in Zhipu’s 2023 funding round.
MiMo’s licensing is plainer than several peers’. Kimi K2 adds a display rule above set user or revenue levels, MiniMax M2.7 restricts commercial use, and GLM-5.3 adds a security review for very large model-hosting businesses; the modified MIT page has the exact wording. MiMo’s open weights have stayed on standard MIT, though some launches, such as V2-Pro, were API-only first.
FAQ
What does MiMo stand for?
Our sources do not say, so we do not guess. It is the name of Xiaomi’s language model family.
Is Xiaomi MiMo open source?
Most of it. The MiMo-7B, V2-Flash, V2.5-Pro and V2.6 weights we checked are listed as MIT on their model cards. V2-Pro, V2-Omni and V2-TTS were API-only.
Can I use MiMo commercially?
MIT permits it. For the 9B distilled model, check the Qwen base model’s licence too, which is Apache 2.0 on its card.
Where do I run MiMo?
Xiaomi offers an API platform and a MiMo Studio, and the weights are on Hugging Face under the XiaomiMiMo organisation.