OpenModelsHQ

Xiaomi MiMo

Beijing, China · Xiaomi · China Desk

Xiaomi MiMo is based in Beijing, China and is part of Xiaomi. We track 15 open-weight models from it, and the newest is MiMo-V2.6-Flash-MOPD, released 27 Sep 2026.

Models tracked
15
LLM 10 · Vision 4 · Speech 1
Newest release
MiMo-V2.6-Flash-MOPD
27 Sep 2026
Country
China
Releases timeline
15 Xiaomi MiMo releases from 29 Apr 2025 to 27 Sep 2026. The largest is MiMo-V2.6-Pro-MOPD at 1.02T parameters. Dot area follows total parameters; a ring means the size is not published.
Releases timeline: data
ModelCategoryReleasedParameters
MiMo-V2.6-Flash-MOPDLLM27 Sep 2026311B
MiMo-V2.6-Pro-MOPDLLM27 Sep 20261.02T
MiMo-V2.6-Distill-Qwen-9BVision21 Sep 20269.4B
MiMo-V2.6-Flash-RLLLM21 Sep 2026311B
MiMo-V2.6-Pro-RLLLM21 Sep 20261.02T
MiMo-V2.5LLM27 Apr 2026311B
MiMo-V2.5-ProLLM27 Apr 20261.02T
MiMo-V2.5-ASRSpeech23 Apr 20267.6B
MiMo-V2-FlashLLM16 Dec 2025310B
MiMo-Embodied-7BVision19 Nov 20258.3B
MiMo-VL-7B-RL-2508Vision7 Aug 20258.3B
MiMo-VL-7B-SFT-2508Vision7 Aug 20258.3B
MiMo-7B-RLLLM29 Apr 20257.8B
MiMo-7B-BaseLLM29 Apr 20257.8B
MiMo-7B-SFTLLM29 Apr 20257.8B

Origin

Xiaomi, the Beijing consumer electronics company best known for smartphones, software and electric vehicles, moved into large-scale AI models in April 2025 when it introduced MiMo, according to Wikipedia. Lei Jun founded Xiaomi in 2010 together with six associates. Its AI model team publishes as Xiaomi MiMo, and the model cards credit the “Xiaomi MiMo Team”.

In March 2026 Reuters reported that a trillion-parameter model circulating anonymously was Xiaomi’s MiMo-V2-Pro, and Lei Jun said the company planned to invest at least US$8.7 billion in AI over the following three years.

How it got here

MiMo-7B came first, in April 2025. Wikipedia describes it as a reasoning model, and the Hugging Face page lists base, SFT and RL versions. MiMo-VL-7B, a vision-language model, arrived in May 2025, with an updated version in August 2025. MiMo-Audio-7B-Instruct appeared in September 2025.

MiMo-V2-Flash came in December 2025. Its model card describes a mixture-of-experts language model with 309 billion total and 15 billion active parameters, designed for fast reasoning and agent workflows. MiMo-V2.5 and MiMo-V2.5-Pro came next, and the Pro card describes an open-source mixture-of-experts model with 1.02 trillion total and 42 billion active parameters.

Then came the MiMo-V2.6 releases: Pro and Flash versions on the Hugging Face page, and a 9B model called MiMo-V2.6-Distill-Qwen-9B. Its card says it was made by fine-tuning Qwen3.5-9B on data generated by MiMo, so it builds on another lab’s open model.

Licensing pattern

Every MiMo model we track uses the MIT licence, from the 7B reasoning model to the trillion-parameter V2.5-Pro. The model 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 it. We found no change of licence between releases.

The 9B model needs a second check. The MiMo-V2.6-Distill-Qwen-9B card lists MIT and names Qwen/Qwen3.5-9B as its base model with a “finetune” relation, so check the base model’s licence as well.

Release lineage

Newest first
  1. MiMo-V2.6-Flash-MOPD311B
  2. MiMo-V2.6-Pro-MOPD1.02T
  3. MiMo-V2.6-Distill-Qwen-9B9.4B
  4. MiMo-V2.6-Flash-RL311B
  5. MiMo-V2.6-Pro-RL1.02T
  6. MiMo-V2.5311B
  7. MiMo-V2.5-Pro1.02T
  8. MiMo-V2.5-ASR7.6B
  9. MiMo-V2-Flash310B
  10. MiMo-Embodied-7B8.3B
  11. MiMo-VL-7B-RL-25088.3B
  12. MiMo-VL-7B-SFT-25088.3B

Largest first. The last line on each tile is the memory of the smallest device that fits the model at 4-bit.