The Xiaomi MiMo team has officially launched and made the new-generation MiMo V2.6 series models entirely open-source. This series includes three versions: Pro, Flash, and Ultraspeed. Notably, the Pro version has soared to the pinnacle of the AA Intelligence Index, achieving a remarkable score of 46. This accomplishment places it ahead of both global open-source models and domestic counterparts. Its performance rivals that of the world's foremost closed-source flagship models, yet its cost is almost 30 times lower than that of Grok 4.7 for an equivalent score, thus establishing a new standard for the "intelligence-to-cost" ratio.

The three models are tailored for distinct scenarios: flagship performance, high-frequency large-scale tasks, and high real-time interaction environments. They excel in domains such as game development, 3D modeling, office document creation, and new material research, focusing on practical applications rather than merely chasing benchmark scores.

Without modifying the pre-training foundation or parameters, this series enhances intelligence through post-training techniques that amplify RL (Reinforcement Learning) computing power. It adopts the innovative MixRL hybrid task reinforcement learning approach, which leverages a self-improving system with three-axis coordination: Rollout computing, environment and Harness, and Grader computing, all supported by stability infrastructure. This series has accomplished the largest single RL training endeavor to date in the open-source realm, with the entire training process transparently disclosed. This showcases RL's exceptional sample efficiency and generalization capabilities.

Furthermore, this series upholds MiMo's commitment to technical efficiency, doubling intelligence while keeping prices constant. It slashes the cost of accessing global cutting-edge intelligence to around $0.1, offering substantial cost-effectiveness benefits.