Xiaomi's MiMo-V2.6-Pro tops open-weight models

Xiaomi released and fully open-sourced the MiMo-V2.6 series: the flagship MiMo-V2.6-Pro and the efficient MiMo-V2.6-Flash, two natively omnimodal models, plus a Pro-UltraSpeed mode that generates at up to 20x the speed. MiMo-V2.6-Pro scores 46.32 on the Artificial Analysis Intelligence Index, the highest of any open-weight model, surpassing Kimi K3 and Qwen3.8 Max while still trailing closed-source leaders like Claude Fable 5.1 and GPT-6 Astra.

What’s actually new

• Scaling RL in public. Xiaomi streamed production RL training live: in under six days both models completed 30 RL steps over roughly 750,000 trajectories, at a cost of about $0.85M (Flash) and $2.62M (Pro). Held-out software engineering scores on DeepSWE v1.1 rose by about 17 points (Flash) and 14 points (Pro).
• Open from top to bottom. Weights, the full technical report, 7,000+ RL task environments, and the end-to-end RL training framework are all on Hugging Face, alongside the new MiMo-V2.6-Distill-Qwen-9B distilled model.
• Agent-grade performance. On most agent benchmarks Pro performs on par with Claude Opus 5 and GPT-5.6 Sol, and it matches them on the Design Arena leaderboard. New capabilities reach 3D scene building, Blender modeling, embodied robot control, computer use, video creation, and music composition.

Why it matters

Xiaomi chose the RSI path: scale reinforcement learning on verifiable tasks, then publish everything needed to reproduce it. The result is an open model at the top of the open-weight ranking with training that was literally watched live. That is a clear signal on where open-source frontier work is heading — measurable, reproducible RL post-training rather than closed runs — even as the strongest proprietary models still hold the overall lead.

(Original source: Xiaomi MiMo)