MiMo v2.6
An iteration of MiMo, the large language model series developed by Xiaomi. The series is designed to serve as the AI capability foundation for Xiaomi's terminal ecosystem, including smartphones, IoT devices, and vehicles. The version emphasizes on-device deployment and inference efficiency.
Core Facts
Timeline (last 90 days)
Xiaomi open-sourced thousands of RL environments along with MiMo v2.6.
The root cause of the tool-call repetition issue in Xiaomi MiMo v2.6 is a reward blind spot in large-scale reinforcement learning—the repetition penalty only takes effect when tool calls exceed 32 times per round.
Having the MiMo model create a product launch landing page in the DeepSeek harness consumed nearly 45 million tokens, of which approximately 98% were cache hits.
MiMo v2.6 can drive Blender to generate 3D assets and animate them with code, but its spatial reasoning is not strong, and in some scenarios for procedurally generating 3D scenes with three.js it underperforms DeepSeek 4.1 Flash.
MiMo models can use Lean to solve mathematical formulas without domain-specific RL, and provide assistance in fields such as materials science and biochemistry.
Xiaomi officially disclosed that the total training cost of MiMo-V2.6 was only $3.5 million.
After MiMo v2.6 was released, it received 293 upvotes and more than 133 comments on Hacker News in a short period of time.
The technical white paper and complete benchmark data for MiMo v2.6 have not yet been publicly disclosed.
The live-stream page simultaneously trains two models, MiMo-V2.6 Pro and MiMo-V2.6 Flash.
Xiaomi has made the post-training process of its Mimo 2.6 model publicly visible through a real-time dashboard.
All Facts (6)
Xiaomi officially disclosed that the total training cost of MiMo-V2.6 was only $3.5 million.
80%VerifiedThe live-stream page simultaneously trains two models, MiMo-V2.6 Pro and MiMo-V2.6 Flash.
65%UnverifiedThe root cause of the tool-call repetition issue in Xiaomi MiMo v2.6 is a reward blind spot in large-scale reinforcement learning—the repetition penalty only takes effect when tool calls exceed 32 times per round.
50%UnverifiedMiMo models can use Lean to solve mathematical formulas without domain-specific RL, and provide assistance in fields such as materials science and biochemistry.
50%UnverifiedXiaomi open-sourced thousands of RL environments along with MiMo v2.6.
50%UnverifiedXiaomi has made the post-training process of its Mimo 2.6 model publicly visible through a real-time dashboard.
50%