For years, the race to build humanity's most powerful AI models was understood to be a contest between a handful of tech giants — Anthropic, OpenAI, Google, Deepseek, Bytedance and Alibaba — with bottomless budgets and decades of institutional knowledge. Then Xiaomi, a Chinese phone maker, barely two years into its AI research division, quietly listed an anonymous model on the world's largest API aggregation platform and watched it top the daily usage charts.
That model was codenamed Hunter Alpha. It is, as Xiaomi confirmed today, an early internal build of MiMo-V2-Pro.
The reveal is significant not only for what MiMo-V2-Pro can do, but for what it represents. The model — a trillion-parameter, 42-billion-active-parameter architecture with a one-million-token context window is definitely among tier 1.5 - tier 2 language models, based on internal feedbacks from CTOL Digital Solutions engineers. More importantly, on the OpenClaw agent framework's standard evaluations, PinchBench and ClawEval, it places third worldwide, trailing only Claude Sonnet 4.6 and Opus 4.6.
Those are not the numbers anyone predicted from a team founded in 2024.
During the Hunter Alpha test phase — before Xiaomi disclosed its identity — the model accumulated over one trillion tokens of usage on OpenRouter. Developers building coding tools drove the majority of that volume. Community testers reported that in direct head-to-head comparisons, Hunter Alpha frequently outperformed Claude Sonnet 4.6. One tester ranked it explicitly between Anthropic's Opus 4.5 and Opus 4.6. Another described it as "a solid first-tier choice."
These assessments came with caveats. A bug in OpenRouter's implementation initially suppressed the model's performance, and a subsequent fix produced what one prolific tester, identified online as van dark, called a dramatic improvement. Post-fix results showed MiMo-V2-Pro completing a Monopoly game implementation "completely correctly, fully functional," building a 3D Dream of the Red Chamber interactive scene that matched the previous best domestic model, and handling code editing tasks that one tester called "almost equivalent to Opus 4.5" — with one particularly complex modification judged to exceed it.
Limitations were openly acknowledged, even by supporters. The model is text-only, despite its trillion-parameter scale. It produces verbose output that hurts practical usability. Its world knowledge trails Gemini and other models with deeper multilingual alignment work. One tester noted that the stated one-million-token context window appears to be a 256,000-token architecture forcibly extended, without proportional attention improvements.
The honest consensus: not frontier-class, but knocking on the door.
What may matter more than the model itself is the economics surrounding it. Xiaomi is pricing MiMo-V2-Pro at one dollar per million input tokens and three dollars per million output tokens for standard contexts — a fraction of Claude Opus 4.6's, and below even Claude Sonnet 4.6's. In China's domestic market, it enters a landscape where GLM-5 and GLM-4.7 have already driven prices to levels that would have seemed implausible months ago.
The commoditization is accelerating. A company that simultaneously produces 3nm chips, an electric vehicle that set a Nürburgring production-car lap record, factory-ready humanoid robots, and now a globally competitive foundation model — all on an annual R&D budget estimated between twenty and thirty billion RMB — is not operating by the logic the industry assumed.
"China's massive AI talent density," one community observer wrote during the Hunter Alpha period, capturing a sentiment that recurred across dozens of independent testers. "So many Chinese companies can catch up with the top LLM companies now."
MiMo-V2-Pro is not the model that dethrones Claude or GPT. Not yet. But it is the clearest evidence to date that the skills, capital, and infrastructure required to build serious frontier AI are no longer concentrated in a few zip codes in California. The benchmarks are converging. The prices are collapsing. And the teams doing it are, in some cases, less than two years old.
The frontier is flattening. The question now is not whether China's AI industry catches up — it is how the incumbents respond when it does.
not investment advice
Sources: https://mimo.xiaomi.com/mimo-v2-pro