Xiaomi Releases Latest Large Model, Pro Version Tops Global Open-Source Rankings!
SOURCE: Zhongguancun Science City
TIME: 2026.09.24
On September 22, Xiaomi officially released and open-sourced its all-new generation MiMo-V2.6 series large models. According to the comprehensive intelligence index rankings of Artificial Analysis, a globally authoritative large model evaluation platform, Xiaomi MiMo-V2.6-Pro topped the global open-source large model rankings with 46 points, ranking first among domestic models. 
MiMo-V2.6 Series Offers Three Product Tiers
MiMo-V2.6-Pro is the flagship reasoning model, omnimodal, ultra-high performance, covering complex projects, long-range tasks, high-value work, cybersecurity, and scientific research needs. MiMo-V2.6-Flash is a omnimodal, high-intelligence, low-cost efficient reasoning model, suitable for professional office high-frequency calls and large-scale tasks. MiMo-V2.6-Pro-Ultraspeed balances flagship performance, providing up to 10 times the reasoning speed, serving strong real-time interactive scenarios. 
Among these, MiMo-V2.6-Pro is the first to support omnimodal input including text, images, video, and audio, becoming the only omnimodal Pro model in the global open-source model lineup. Both base models in the MiMo-V2.6 series support omnimodal input, filling the gap in open-source Pro-level models. Targeting real professional workflows, the series focuses on covering complex task reasoning, multi-Agent collaboration, large-scale programming projects, as well as Office, web and front-end interaction design, video editing, 3D generation, music creation, computer control, cybersecurity, scientific research, and other task types. 
In terms of pricing, the MiMo-V2.6 series has been launched on the Xiaomi MiMo Open Platform and the Xiaomi MiMo Desktop client, and continues to use the API pricing of the V2.5 series. Artificial Analysis intelligence index single-task cost ranking data shows that Xiaomi MiMo-V2.6-Pro’s actual measured single-task cost is USD 0.13, approximately less than CNY 1, significantly lowering the barrier to using top-tier AI. 
It is reported that MiMo-V2.6 is one of the largest explorations of reinforcement learning training scale among current domestic open-source models. The average pass rates of the two models on training tasks have improved by 25% and 12% respectively. Training has been intensified from three aspects: first, larger batch sizes and higher throughput, with 1,568 samples used per update, supporting 1M context length training, and single-step training tokens reaching 3.5–3.7B; second, more tasks and complex environments, covering Code, General, Visual, Cyber, and other directions; third, larger Grader computing power, enabling the model to complete tasks with shorter paths and fewer tokens.
 
Additionally, the MiMo-V2.6 series also possesses certain capabilities in scientific research and cybersecurity offense and defense. The series has been applied to Xiaomi’s internal new materials R&D work, proposing design schemes for several metal-organic frameworks (MOF) materials. It can retrieve and organize literature and patents, propose research hypotheses, evaluate the novelty of design schemes, and conduct “dry experiments” by automatically building simulation environments, calculating binding strength between materials and PFAS, and screening the most promising candidate materials for subsequent “wet experiment” verification. Xiaomi stated that it has assisted researchers in improving work efficiency by over 20 times, with the usability of related achievements significantly surpassing industry peers, and has received preliminary recognition from the academic community. 
返回上一页