Xiaomi Open-Sources MiMo-V2.6 Omni-Modal Models, Tops Open-Weight Rankings on AA Index

Xiaomi has officially released and open-sourced the Xiaomi MiMo-V2.6 series, including two native omni-modal models: Pro and Flash. Xiaomi calls this a key step in exploring recursive self-improvement, scaling up reinforcement learning compute to push the boundaries of intelligence through continuous exploration and feedback.

MiMo-V2.6-Pro scored 46 on the Artificial Analysis Intelligence Index v4.3.2, beating Kimi K3 and GLM-5.3 to become the highest-ranked open-weight model on the AA index, up 20 points from its predecessor MiMo-V2.5-Pro's 26. Xiaomi also noted the model still lags behind top closed-source models like Claude Fable 5.1 and GPT-6 Astra, which both scored 53.

Training took under 6 days for both Pro and Flash, costing about $2.62 million and $850,000 respectively, with 30 steps each and roughly 750,000 trajectories. The model combines 3D spatial reasoning, multimodal perception and computer operation capabilities. It can control a Franka Panda robotic arm for grasping and placement, assist with MOF material screening, and complete Lean 4 formalization of mathematical theorems.

MiMo Desktop client and membership plans are now available. The series is live on the Xiaomi MiMo open platform, with model weights and technical reports fully open-sourced.