On September 22, 2026, Xiaomi released MiMo-V2.6, a new open-source model family that enters the top tier of publicly available AI — and undercuts the pricing of every proprietary rival by an order of magnitude.
What's in the family
The release includes three models:
- MiMo-V2.6-Pro: 1.02 trillion parameters, 42B active (sparse mixture-of-experts), 1M-token context
- MiMo-V2.6-Flash: 309B parameters, 15B active, same 1M context, optimized for throughput
- MiMo-V2.6-Pro-UltraSpeed: Delivers up to 20× faster output than Pro at comparable quality
All three are natively omnimodal — they process text, image, video, and audio in a single model with no glued-together adapter pipeline. The vision encoder carries 681 million parameters; audio handling runs through a 308M-parameter AudioTokenizer and 127M-parameter audio patch encoder. All weights are available under an MIT license on Hugging Face.
Benchmarks
MiMo-V2.6-Pro scores 46.32 on the Artificial Analysis Intelligence Index, making it the highest-ranked open-weights model at launch, ahead of Kimi K3 (44) and GLM-5.3 (45). On AutomationBench, it reaches 53.1 versus Claude Opus 5's 50.3. On Agents' Last Exam, it ties Opus 5 at 31.6. Terminal Bench 2.1: 89.9 versus Opus 5's 89.1.
These numbers sit at or above the current proprietary frontier on agentic and reasoning benchmarks — while the model is fully open and self-hostable.
The price gap
Xiaomi's API pricing: Flash at $0.14/$0.28 per million tokens; Pro at $0.435/$0.87; Pro-UltraSpeed at $4.35/$8.70. Compared against GPT-6 Astra and Claude Fable 5.1, which run $10–$50 per million tokens, Xiaomi claims Pro is "one-20th to one-60th the cost of overseas models at comparable intelligence." Even at the high end, that math significantly changes the economics for any team doing high-volume inference.
Who it's for
Xiaomi highlighted three use-case categories at launch: game world building and 3D modeling (where combined image, text, and spatial reasoning matter), embodied simulation, and multimodal business workflows that mix documents, audio transcripts, screenshots, and logs in a single context. The MiMo AI Studio and MiMo Desktop apps provide hosted access; the model is also available on OpenRouter and MiMo Code for developer integrations.
For organizations that have been waiting for an open-weights model that competes with the proprietary frontier on multimodal agentic tasks, MiMo-V2.6 is the clearest answer yet.