Xiaomi's Secret 1T-Param Model Stress-Tested Itself on the Open Web

On March 11, a 1-trillion-parameter model called Hunter Alpha appeared on OpenRouter with zero attribution — no company, no press release, no social posts. It described itself only as a Chinese AI with a 1M-token context window, processed 160 billion tokens in its first week, and topped the platform's usage leaderboard before anyone knew who built it.

The internet blamed DeepSeek V4 — the training cutoff matched, the reasoning patterns looked familiar. Others pointed to Zhipu AI based on behavioral fingerprints. Both were wrong. On March 18, Xiaomi confirmed it was an early build of MiMo-V2-Pro, an agentic model led by former DeepSeek-V2 engineer Luo Fuli — which explained the stylistic resemblance.

MiMo-V2-Pro sits 10th globally on the Artificial Analysis Intelligence Index and ranks as the highest-rated Chinese model on GDPval-AA Elo. It costs roughly one-seventh of comparable OpenAI/Anthropic models.

Why it matters: Xiaomi ran a live, global stress test of a frontier model with no announcement — and the model passed. The age of stealth launches is here, and attribution is now a lagging indicator.