Xiaomi MiMo-V2-Pro: The Hunter Alpha Reveal

On March 11, 2026, a model called "Hunter Alpha" appeared anonymously on OpenRouter with no lab name, no model card, and a 1-trillion-parameter budget. Within a week it had processed over a trillion tokens, topped OpenRouter's daily usage charts, climbed to #8 globally on the Artificial Analysis Intelligence Index, and convinced most of the local-LLM community it was an unannounced DeepSeek V4. On March 18, Xiaomi's MiMo team stepped forward and claimed it as MiMo-V2-Pro. The technical story is interesting. The bigger story is that Xiaomi just demonstrated it can stealth-launch a frontier model and have it taken seriously without leaning on its brand.

The seven days nobody knew who built it

Hunter Alpha showed up on OpenRouter on a Wednesday with a free tier and no documentation. The first thing the community noticed was that the system prompt identified it as "a Chinese AI model primarily trained in Chinese" with a May 2025 training cutoff, which matched DeepSeek's published cutoff exactly. Reverse-engineered parameter counts pegged it at roughly 1T total, again consistent with leaked V4 specs that had been circulating since February. A community write-up on Medium walked through the evidence and assigned DeepSeek as the most likely author.

The benchmarks did not slow the speculation down. SWE-bench Verified scores landed near 78%. ClawEval came in at 61.5, putting it third worldwide behind Claude Opus 4.6 (66.3). PinchBench numbers were class-leading. The token volume on OpenRouter spoke even louder than the leaderboards: developers were paying for the calls and the dominant call patterns were coding agents and tool-use loops.

The reveal

On March 18, Luo Fuli, head of Xiaomi's MiMo division, claimed Hunter Alpha as an early internal build of MiMo-V2-Pro. The MiMo team had been a quiet presence on HuggingFace for months: MiMo-7B (MIT-licensed, mobile-class), MiMo-VL-7B for vision-language, MiMo-Audio-7B, and a 309B-total / 15B-active MoE called MiMo-V2-Flash in December 2025. None of those releases got significant Western attention. The 1T MiMo-V2-Pro was a sharp step up the scale ladder, and the team's choice to debut it anonymously meant the community formed its opinion before the brand could prejudice the read.

The pricing matters. At $1 / $3 per million tokens, MiMo-V2-Pro is roughly 5 to 7 times cheaper than Claude Opus 4.6 or GPT-5 at comparable benchmark tiers. That is the same cost-down playbook DeepSeek used in late 2024 to reset the API price floor; Xiaomi has just done it again at the frontier.

Why the DeepSeek attribution stuck

Three things made Hunter Alpha read as DeepSeek to the community. First, a "V4 Lite" variant had reportedly appeared on DeepSeek's own site days before, raising expectations. Second, the parameter count, context length, and training cutoff all matched leaked V4 specs. Third, the model's strengths (structured agentic workflows, tool-calling, code) and weaknesses (creative writing, hard math) lined up with DeepSeek's known training emphasis. A handful of analysts flagged GLM-6 from Zhipu AI as an alternative theory, since the same OpenRouter account had previously hosted GLM-5 as "Pony Alpha." Almost nobody guessed Xiaomi.

Why Xiaomi at all

Xiaomi's prior AI work is the missing context. MiLM-6B landed in August 2023 as the company's first public LLM. MiLM2-30B followed as a cloud-scale internal model powering HyperOS features. HyperOS 3.0 in 2026 layered visible AI features into the phone and IoT line. Then, in late 2025, Luo Fuli joined Xiaomi from DeepSeek, where she had been a key contributor to V2. That hire was the visible signal that Xiaomi intended to compete at frontier scale, not just at the on-device tier.

MiMo-V2-Pro is the first product of that pivot. A faster follow-up, MiMo-V2.5-Pro, landed on April 22, 2026 and matched frontier benchmarks at lower token cost, suggesting the team is iterating quickly and is not running on one-shot luck.

The stealth-launch pattern

Anonymous OpenRouter releases are not new. Anthropic's "claude-3-opus-pre" leaked briefly in early 2024. Several smaller Chinese labs have used the same channel to dogfood unreleased models against real developer traffic. What makes the Xiaomi case different is the scale and the credibility: a 1T-parameter frontier candidate, deployed to production-grade load on an open marketplace, then claimed only after a week of independent benchmarks had already validated it.

For practitioners, the lesson is not "Xiaomi is now in the top tier" (one frontier model does not equal a moat). The lesson is that the route from "we have a frontier model in the lab" to "the community treats it as real" can be one week and one OpenRouter account, with the lab name attached at the end of the validation cycle instead of the start. That changes how to read leaderboards. The next time an anonymous OpenRouter listing posts top-tier benchmark numbers, the default assumption should no longer be "leaked Anthropic or OpenAI checkpoint." It should be "another lab is testing the water."

What to do with it today

MiMo-V2-Pro is callable on OpenRouter right now at the listed pricing. The model is API-only, no open weights, so this is not one for the local-LLM stack. If you build agent or coding workflows that already use Claude Opus or GPT-5 through OpenRouter, MiMo-V2-Pro is a straight drop-in to test against. Run it on your actual evals before forming an opinion. The benchmark profile says it will be excellent at structured tool-use loops and merely OK at long-form creative work; verify whether that matches your traffic.

And keep half an eye on the OpenRouter homepage. The next anonymous 1T-parameter listing might be from Tencent, Bytedance, Baichuan, or someone you have not heard of yet. The stealth-launch playbook now has a working precedent.

Sources and further reading

• VentureBeat: Xiaomi stuns with MiMo-V2-Pro
• TopAIProduct: 1T-parameter mystery model
• Community analysis of the anonymous week
• Apiyi MiMo-V2-Pro API guide and full benchmarks
• MarkTechPost on the V2.5 follow-up
• DIGITIMES on the Luo Fuli hire

Timeline and benchmark numbers compiled from the linked sources. Not independently verified. Compiled 2026-05-19.