On September 22, 2026, Xiaomi officially released and open-sourced the MiMo-V2.6 series. The Pro version scored 46 points on the Artificial Analysis Intelligence Index v4.3.2, surpassing Kimi K3's 44 points and GLM-5.3's 45 points, making it the highest-ranked open-weight model on the current AA Index.
The Facts
Xiaomi states that MiMo-V2.6-Pro improved by 20 points over its predecessor MiMo-V2.5-Pro's 26 points, while still trailing closed-source models such as Claude Fable 5.1 and GPT-6 Astra, which sit at 53 points. Training Pro and Flash took less than 6 days, costing about $2.62 million and $850,000 respectively, with each completing 30 steps and roughly 750,000 trajectories in total. Average pass rates on training tasks rose by 25% and 12% respectively on a relative basis, and on the out-of-sample long-horizon software engineering benchmark DeepSWE v1.1 they improved by about 17 points (48.8 to 65.7) and about 14 points (58.4 to 72.6) respectively.
This training scaled RL compute along three dimensions: larger batches and higher throughput, with 1,568 samples used per update, support for training with a 1 million context length, and 3.5 to 3.7B tokens per training step; more tasks and more complex environments, building a multi-task training system spanning Code, General, Visual, Cyber and other directions; and greater grader compute, delivering more precise reward signals for long-horizon RL tasks through relative comparison within groups.
How It Works
The MiMo-V2.6 series integrates 3D spatial reasoning, multimodal perception and computer operation capabilities. In game development scenarios, after a user inputs images, video or text, the model can break the requirements down into multiple tasks, with multiple agents collaborating to build 3D scenes, write interaction logic and perform visual checks. In embodied simulation environments, the model can take multi-view camera footage directly as input and, through a closed visual-feedback loop, control a Franka Panda robotic arm to grasp objects, match colors and place them precisely.
On the research side, MiMo-V2.6-Pro helped screen design candidates for metal-organic framework materials used to adsorb per- and polyfluoroalkyl substances, and formalized the original main theorem of the classic Li-Yorke paper "Period Three Implies Chaos" in Lean 4, producing more than 6,000 lines of Lean source code whose complete proof passed the Lean kernel's check. The model received no specialized post-training for Lean.
Industry Impact
The MiMo-V2.6 series follows the API pricing of the V2.5 series: Flash costs 1 yuan per million tokens for input and 2 yuan for output; Pro costs 3 yuan and 6 yuan, with a 99% cache discount. At equivalent levels of intelligence, the price is only 1/20 to 1/60 that of overseas models. The MiMo Desktop client launched at the same time, supporting Windows and Mac; a membership subscription gives access to the Pro and Flash models, and users can also configure their own API key.
Xiaomi has fully open-sourced the MiMo-V2.6-Pro and Flash model weights and technical report, and is also releasing MiMo-V2.6-Distill-Qwen-9B along with supporting RL research resources, including 7k+ high-quality RL task environments, an end-to-end RL training framework and a lightweight, composable Harness. These resources cover four categories of agent tasks: software engineering, vulnerability reproduction, knowledge work, and web design and development.
Strategic Assessment
(The following is analysis rather than fact.) Through large-scale investment in RL compute, the MiMo-V2.6 series has carved out a lead among open-source models, but a clear gap with top-tier closed-source models remains. The simultaneous release of open weights and training resources may lower the barrier for the community to reproduce results and speed up iteration in multi-task environments. The price advantage and the launch of a desktop client may create some competitive pressure for existing API service providers, but the actual scale of user migration remains to be seen.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接