Xiaomi / VentureBeat / Hugging Face
Important
XiaomiOpen SourceMiMoChina AIXiaomi open-sources MiMo-V2.6: top open-weight model on Artificial Analysis index
September 22, 20262 min read
Xiaomi released and open-sourced MiMo-V2.6-Pro and MiMo-V2.6-Flash under MIT. Pro scores 46 on the Artificial Analysis Intelligence Index, surpassing Kimi K3 and Qwen3.8 Max to become the strongest open-weight model. Both are natively omnimodal with 1M-token context.
Why it matters
A phone-and-EV company just shipped the leading open-weight frontier model and released full RL training resources. Reinforces that Chinese labs are closing the open-weight gap with large-scale reinforcement learning.
Xiaomi released the MiMo-V2.6 series on September 22 and open-sourced both MiMo-V2.6-Pro and MiMo-V2.6-Flash under the MIT license. Pro scores 46.32 on the Artificial Analysis Intelligence Index, ahead of Kimi K3 and Qwen3.8 Max and making it the strongest open-weight model on that leaderboard. It still trails the top closed models (Claude Fable 5.1 and GPT-6 Astra at ~53).
Both models are natively omnimodal with a claimed 1-million-token context. Flash is a 309B total / 15B active MoE; Pro is larger (reported ~1T total / 42B active). Xiaomi says the models underwent one of the largest RL compute runs among Chinese open-weight efforts, completing multi-task RL in under six days. The company also released technical reports, 7,000+ RL task environments, training frameworks, and a distilled 9B model.
Weights are on Hugging Face. API access and MiMo Desktop are live. The release is a clear signal that open-weight competition is intensifying beyond the usual Chinese labs.