Xiaomi dropped three AI models on March 18: a trillion-parameter LLM, an omni-modal agent model, and a text-to-speech system. The flagship, MiMo-V2-Pro, turns out to be the anonymous "Hunter Alpha" that topped OpenRouter's usage charts last week, racking up over a trillion tokens before anyone knew who built it. Led by Fuli Luo, a former DeepSeek R1 researcher, the MiMo team called the stealth launch a "quiet ambush" on the global frontier.
MiMo-V2-Pro runs a sparse MoE architecture: 1T total parameters, 42B active per forward pass, with a 1M-token context window. On Xiaomi's own benchmarks, it scores 84.0 on PinchBench (third globally, behind both Claude 4.6 variants) and 61.5 on ClawEval, above GPT-5.2's 50.0. Artificial Analysis placed it at 49 on its Intelligence Index. These are strong numbers, though the agentic benchmarks skew toward Xiaomi's own OpenClaw framework. Independent verification beyond Artificial Analysis is limited so far.
The second model, MiMo-V2-Omni (tested anonymously as "Healer Alpha"), fuses text, image, video, and audio through dedicated encoders into a shared backbone. Xiaomi claims it handles continuous audio exceeding 10 hours in a single pass, with audio understanding surpassing Gemini 3 Pro on select benchmarks. Parameter count wasn't disclosed. A TTS model rounds out the trio, supporting multiple Chinese dialects, emotional control, and singing.
Pricing undercuts Western competitors by roughly 5x: MiMo-V2-Pro runs $1/$3 per million tokens (input/output) at up to 256K context, $2/$6 for 256K to 1M. The Omni model is $0.40/$2.00. Developers get one week of free access through agent frameworks including OpenClaw, Cline, and Blackbox. APIs are live at platform.xiaomimimo.com.
Bottom Line
MiMo-V2-Pro scores third globally on agent benchmarks at roughly one-fifth the price of Claude Sonnet 4.6, per Xiaomi's own data.
Quick Facts
- 1T total parameters, 42B active (MoE architecture)
- 1M-token context window
- PinchBench: 84.0 (3rd globally, company-reported)
- ClawEval: 61.5 (3rd globally, above GPT-5.2)
- API pricing: $1/$3 per M tokens (up to 256K context)
- MiMo-V2-Omni: $0.40/$2.00 per M tokens
