Xiaomi’s MiMo-V2.6-Pro becomes the top open-weights model — trained for about $2.6M

Xiaomi released three models on September 21: a trillion-parameter Pro, a cheaper Flash, and a distilled 9B. The Pro scores 46 on Artificial Analysis' Intelligence Index — the highest of any open-weight model, ahead of GLM-5.3 at 45 and Kimi K3 at 44, and level with the Grok 4.7 that xAI shipped the same day. The best closed models still sit at 53.

The number that matters is the other one. Xiaomi says Pro cost about $2.62 million to train and Flash about $850,000, in under six days of reinforcement learning across roughly 750,000 trajectories. More than half that budget went on generating and grading the model's own experience before any of it became weights — a method the team calls “You Only RL Once.” Fuli Luo, who leads the MiMo team and came from DeepSeek, framed it as a deliberate bet: “In an era when compute is brutally scarce, we still chose to dedicate a team of several dozen people to one goal: scaling up RL.”

At $0.435 per million input tokens for Pro and $0.14 for Flash, the open tier now undercuts the frontier by an order of magnitude at roughly nine-tenths of the measured capability. That is the pressure every closed lab is pricing against — and it is coming almost entirely out of China.

• Reported Pro scores 46 on Artificial Analysis’ Intelligence Index, level with Grok 4.7 and ahead of Gemini 3.8 Flash (41) and DeepSeek V4.1. VentureBeat
• Reported 1.02 trillion parameters (42B active), 1M-token context; $0.435 in / $0.87 out per million tokens, with Flash at $0.14 / $0.28. VentureBeat
• Claimed Xiaomi says Pro cost about $2.62M to train and Flash about $850,000, in under six days of reinforcement learning. VentureBeat