Xiaomi released MiMo V2 Pro this week, a 1-trillion-parameter model built for agent tasks. You can try it for free right now through MiMo Chat or the MiMo Claw agent.
The model first showed up on OpenRouter on March 11 as “Hunter Alpha,” with people guessing it was from DeepSeek. Turns out Xiaomi built it, led by Luo Fuli, who worked on DeepSeek R1 before. Logos reports it has a 1-million-token context window and supports frameworks like OpenClaw.
Performance Claims
Artificial Analysis put MiMo V2 Pro eighth overall, behind GPT 5.2 and Z.ai’s GLM-5. The same tests gave it tenth on the Global Intelligence Index, top for Chinese models. In GDPval-AA for agent tasks, it scored an Elo of 1,426, beating GLM-5 at 1,406. Digital Today notes its sparse setup only fires 42 billion parameters per inference.
Xiaomi’s internal tests show it beating Claude 4.6 Sonnet on code generation and matching Opus 4.6 overall in agent work. Luo called the launch a “quiet ambush” on X, saying the team shifted hard to agent systems after seeing what they could do. Times of India quotes his full post.
Cost Edge
API access costs about one-sixth to one-seventh of OpenAI or Anthropic models for up to 256,000 tokens, per VentureBeat coverage in those reports. Developers get a free week to test it. Xiaomi plans to open-source a variant once stable.
