Xiaomi Open-Sources MiMo-V2.6 After Large-Scale RL
Xiaomi releases MiMo-V2.6 Pro and Flash with open weights, RL environments and training tools for agent research.
Xiaomi opens the full training stack
Xiaomi has released and open-sourced the MiMo-V2.6 family, consisting of the multimodal MiMo-V2.6-Pro and MiMo-V2.6-Flash. The company says both models were improved through a six-day live reinforcement-learning run, with approximately 750,000 training trajectories and reported costs of about $850,000 for Flash and $2.62 million for Pro.
The release includes model weights, a technical report, more than 7,000 reinforcement-learning environments, an end-to-end training framework, and lightweight agent harnesses. The environments cover software engineering, vulnerability reproduction, knowledge-intensive work, and web design and development. Xiaomi also released MiMo-V2.6-Distill-Qwen-9B and related research resources.
A stronger open-weight contender
Xiaomi reports that MiMo-V2.6-Pro reached 46 on the Artificial Analysis Intelligence Index, making it the highest-ranked open-weight model on that index at launch. It says the model is competitive with leading closed systems on several agent benchmarks, while still trailing the strongest proprietary models overall.
The most notable claim is about improvement during reinforcement learning. Xiaomi says Pro’s out-of-sample DeepSWE v1.1 score rose from 58.4 to 72.6, while Flash increased from 48.8 to 65.7. The smaller distillation release also improved across all 11 listed evaluation metrics after reinforcement-learning training.
Why it matters
The strategic importance is broader than another open model joining the leaderboard. Xiaomi is publishing the environments, harnesses and training components needed for others to reproduce or extend agentic reinforcement learning. That lowers the barrier for laboratories that have capable base models but lack task environments and evaluation infrastructure. The claims still depend heavily on Xiaomi’s own measurements, and the release date difference between the announcement page and public artifacts should be tracked carefully, but the openness of the surrounding stack gives this launch unusual research value.
Uncle Cat take
The 7,000-plus task environments may matter more than the 46-point index score: Xiaomi is exporting the machinery that can make agent training repeatable.