Xiaomi MiMo V2: Three AI Models Built for Agents, Robots, and Voice

Before Xiaomi officially announced MiMo V2 Pro, the model spent weeks on OpenRouter under the codename “Hunter Alpha” — accumulating a trillion tokens and topping daily rankings while most observers attributed the results to DeepSeek.

The reveal punctuates a pattern: Chinese AI labs are not iterating on the margins. They are shipping production-grade models at price points that restructure the economics of the entire stack.

What Xiaomi Launched

MiMo V2 is three distinct models, each targeting a different layer of the agentic infrastructure problem.

MiMo V2 Pro is the flagship language model. Its architecture deploys a Mixture-of-Experts design with one trillion total parameters and 42 billion active per request — roughly three times the scale of Xiaomi’s December 2025 predecessor. On PinchBench, it scores 81.0, placing third globally behind Claude Opus 4.6 and one other model. On SWE-bench Verified — the standard benchmark for autonomous coding — it reaches 78%, against Claude Opus 4.6’s 80.8%. The context window extends to one million tokens via hybrid attention.

The pricing is the differentiator that matters operationally: $1 per million input tokens, $3 per million output. Claude Sonnet 4.6 runs $3–$15; Opus runs $5–$25. For high-throughput agentic pipelines where the token cost is a real constraint, that gap reshapes which models are viable at scale.

MiMo V2 Omni integrates image, video, and audio encoders for autonomous decision-making across modalities. Demonstrated capabilities in the release include independent price comparison and negotiation for online shopping, dashcam hazard detection, multimedia content creation, code debugging, and publication to external platforms — all within a single inference session. The agent benchmark score is 54.8, against Claude Opus 4.6’s 66.3, which signals the multimodal reasoning gap remains real but is closing faster than most trajectory forecasts projected.

MiMo V2 TTS addresses the voice layer. Trained on over 100 million hours of speech data, the model generates voice from natural language descriptions rather than dropdown-style configuration. It handles native paralinguistics — coughs, sighs, hesitation — and interprets typographic signals like capitalization and repetition as emphasis cues. A single model handles both speech and singing synthesis.

The “Hunter Alpha” Signal

The anonymous OpenRouter deployment before launch was not incidental. Running under a mystery codename allowed Xiaomi to accumulate real-world usage data — a trillion tokens — without the benchmark gaming incentives that distort pre-announcement model behavior. That MiMo V2 Pro topped daily rankings while misidentified as DeepSeek is a meaningful data point: the model is competitive on real tasks, not just curated evals.

DeepSeek V4 was delayed during this period due to reported scaling challenges, which sharpened the contrast between the two most-watched Chinese labs.

Content created with AI assistance and reviewed for accuracy.