Overview
- Xiaomi released and open‑sourced the MiMo‑V2.6 family, publishing Pro and Flash weights plus technical reports and an end‑to‑end RL training framework for public use.
- The firm says it used heavy short‑burst RL training to boost performance, reporting about 750,000 trajectories in under six days and estimated costs of roughly $2.62 million for Pro and $0.85 million for Flash.
- MiMo‑V2.6‑Pro scored 46 on the AA index, which Xiaomi presents as the highest among open‑weight models while acknowledging a remaining gap to top closed models that score around 53.
- The models add full‑modality support and 3D spatial reasoning, with demonstrations that include multi‑view control of a Franka Panda arm for grasping and academic use cases such as MOF candidate screening and a Lean formal proof.
- Xiaomi also launched MiMo Desktop with a membership tier and an UltraSpeed inference mode while keeping V2.5 API pricing to encourage developer adoption and community replication of its RL methods.