Xiaomi · 2026-03-18 · major
MiMo-V2-Pro — Xiaomi's trillion-parameter agentic model
Xiaomi's stealth-launched flagship: 1T+ total params (42B active MoE), 1M context, agentic-optimized. Appeared anonymously on OpenRouter as 'Hunter Alpha' before official reveal. $1/M input tokens.
A trillion-parameter model from Xiaomi that appeared anonymously on OpenRouter, got mistaken for DeepSeek V4, and turned out to be a genuine frontier contender at $1/M input tokens.
Key specs
| Parameters | 1T+ |
|---|---|
| Active params | 42B |
| Price | $1/M input |
| Pinch bench | 81.0 (#3 globally) |
What is it?
MiMo-V2-Pro is Xiaomi's flagship AI model, officially launched March 18, 2026, after a week-long anonymous stint on OpenRouter under the codename 'Hunter Alpha.' It has over 1 trillion total parameters with 42 billion active per request via a mixture-of-experts architecture, and supports a 1 million token context window. Xiaomi built it specifically for agentic workflows rather than chat.
How does it work?
The model uses a Hybrid Attention mechanism with a 7:1 hybrid ratio and Multi-Token Prediction for fast generation. It ranks 3rd globally on PinchBench (81.0) and 3rd on ClawEval (61.5). Its coding ability surpasses Claude 4.6 Sonnet, and general agent performance approaches Opus 4.6. Pricing is $1 per million input tokens up to 256K context and $3 per million output tokens.
Why does it matter?
A phone company shipping a frontier-tier foundation model is surprising enough. But the stealth launch strategy — letting the model prove itself anonymously against established competitors before revealing the brand — is a statement about how model quality, not brand, drives adoption on API platforms. At $1/M input tokens, it undercuts Western frontier models by 2-15x.
Who is it for?
API users looking for frontier-quality agentic models at lower cost, teams building coding agents.
Try it
openrouter.ai/xiaomi/mimo-v2-pro