Xiaomi open-sources MiMo-V2.6, an RL-scaled model
Xiaomi has released open-weight MiMo-V2.6 models that it says top all open models on the Artificial Analysis index, along with disclosed RL training costs and the code and environments used.
Xiaomi announced the open-source MiMo-V2.6 series. It has two natively multimodal models, Pro and Flash, and the company describes the release as a step toward recursive self-improvement through scaled-up reinforcement learning.
Xiaomi says MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index, ahead of Kimi K3 and Qwen3.8 Max. It says that makes Pro the strongest open model available. The company also says Pro still trails the leading closed models, Claude Fable 5.1 and GPT-6 Astra. Xiaomi says Pro matches Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks, and that Flash now beats the older MiMo-V2.5-Pro. All benchmark figures come from Xiaomi.
The company published unusually detailed training numbers. Each model ran a "Live RL" phase of under six days with 30 steps, producing about 750,000 trajectories in total. The RL phase cost about $850,000 for Flash and $2.62 million for Pro. Over that phase, Xiaomi reports these gains on DeepSWE v1.1, a software-engineering benchmark the models were not trained on:
- Flash: 48.8 to 65.7
- Pro: 58.4 to 72.6
Each update used 1,568 samples, supported 1M-token contexts and processed 3.5–3.7B tokens per step. Xiaomi froze the MoE router to stop load on the model's experts from drifting. It also added several defenses against reward hacking.
Xiaomi says it has open-sourced the technical report, training environments and RL code. API prices stay the same as V2.5, which the company says comes to 1/20 to 1/60 the cost of overseas models at the same intelligence level. The release also highlights 3D game generation and Blender modeling.