Xiaomi’s MiMo-V2.6-Pro Is Now the Strongest Open-Weight AI Model

Xiaomi has pushed out a new MiMo-V2.6 family and published the weights on Hugging Face under an MIT license. The headline model, MiMo-V2.6-Pro, now sits at 46 points on the Artificial Analysis Intelligence Index, which makes it the highest-rated open-weight model in that ranking. It edges past Z.ai’s GLM-5.3 and Moonshot AI’s Kimi K3, both on 44, and it marks a huge jump from Xiaomi’s own MiMo-V2.5-Pro, which scored 26.

That score still leaves it behind the closed frontier. In the combined open-and-closed ranking, MiMo-V2.6-Pro places sixth, with five proprietary models ahead of it: Claude Fable 5.1 and GPT-6 Astra at 53, Claude Opus 5 at 51, Muse Spark 1.3 at 48, and GPT-5.6 Sol at 47. Xai’s Grok 4.7 is tied with Xiaomi at 46. So the gap is real, but it is no longer the canyon it used to be.

The more interesting part is price. Artificial Analysis pegs MiMo-V2.6-Pro at about 0.11 euros per Intelligence Index task, which puts it on the Pareto frontier for intelligence versus cost. It also runs at roughly 125 output tokens per second, good for 12th place among 114 tested models. The catch is verbosity: the benchmark run burned 140 million output tokens.

Technically, this is a sparse mixture-of-experts model with 1.02 trillion total parameters, though only 42 billion are active per inference. Xiaomi says both Pro and the smaller Flash variant handle text, image, speech, and video as input, generate text, and support a one-million-token context window. API pricing stays at the old level, and the UltraSpeed version costs ten times as much as standard Pro.

Xiaomi is also making a point of showing its work. It says the leap came from scaled reinforcement learning on verifiable tasks, with the production run streamed live, and it is releasing the technical report, training environments, and RL code. The model’s strongest areas are pretty broad: cybersecurity, workflow automation, tool use, visual coding, software engineering, and even some research demos involving 3D scenes, a robotic arm, a metal-organic framework for PFAS binding, and a Lean 4 proof spanning more than 6,000 lines. But it still struggles in long terminal sessions, where the proprietary models keep a clear lead.