Summary: At a launch event in Beijing on March 18, 2026, Xiaomi rolled out three in-house models in one stroke: the flagship MiMo-V2-Pro, the omni-modal MiMo-V2-Omni, and MiMo-V2-TTS, which turns text into speech. For developers the takeaway is practical: the unexplained but strong performers near the top of OpenRouter’s charts deserve real scrutiny — any of them could be someone’s next flagship, and prices during the anonymous stretch tend to be set aggressively low.
At a launch event in Beijing on March 18, 2026, Xiaomi rolled out three in-house models in one stroke: the flagship MiMo-V2-Pro, the omni-modal MiMo-V2-Omni, and MiMo-V2-TTS, which turns text into speech. Reuters reported that the very next day, CEO Lei Jun pledged to put at least $8.7 billion into AI over the following three years. A company known for phones and home appliances has now joined the trillion-parameter race.
The money alone is not what makes this launch remarkable. MiMo-V2-Pro carries more than 1 trillion parameters in total, activates 42 billion of them per inference, and reads a 1-million-token context window — figures that put it in the same weight class as frontier-lab flagships. Lei Jun walked through the technical details himself at the Beijing event. How the model first showed up was the theatrical part: ahead of the announcement, it sat on OpenRouter under the alias “Hunter Alpha,” held the top of the daily usage rankings for days on end, and handled over a trillion tokens in total; users guessed it was DeepSeek’s new system until Xiaomi identified it as an early internal test build of MiMo-V2-Pro. Beside it, the omni-modal MiMo-V2-Omni takes in image, video, audio, and text within a single model, while MiMo-V2-TTS completes the set with speech generation.
The $8.7 Billion Commitment
Per Reuters, Lei Jun made the pledge on March 19: at least $8.7 billion for AI spread across three years. Place that beside Xiaomi’s standing research budget of roughly 30 billion yuan a year and the reallocation reads clearly — models, compute, and applications are being advanced at the same time. Running the model team is Luo Fuli, a former DeepSeek researcher who came to Xiaomi in late 2025. Recruit a DeepSeek lead first, then bankroll the infrastructure — it is a playbook with proven results, and Xiaomi is following it.
Testing in Disguise: Hunter Alpha
Debuting new models under an alias is turning into the industry’s standard routine: publish with a codename, soak up genuine traffic without the brand halo or the weight of public opinion, then unmask once the numbers hold up. MiMo-V2-Omni ran the same play as “Healer Alpha.” The trick pays the lab twice: a stretch of anonymous chart-topping is demand evidence no internal benchmark can produce, and it cost Xiaomi nothing beyond inference. For developers the takeaway is practical: the unexplained but strong performers near the top of OpenRouter’s charts deserve real scrutiny — any of them could be someone’s next flagship, and prices during the anonymous stretch tend to be set aggressively low.
1T+ Total, 42B per Inference
MiMo-V2-Pro runs on a mixture-of-experts architecture: a parameter pool beyond 1 trillion with 42B switched on for each inference, and a context that stretches to 1 million tokens. The balance is aimed at agentic workloads — long documents, strings of tool calls, tasks that cross sessions — and the official positioning states outright that it is built to be “the brain of AI agents.” Independent evaluator Artificial Analysis has already ranked it among the front-runners on its intelligence index, and free trials were stretched to April 2 through partners such as OpenClaw after demand proved heavier than capacity.
Closed Today, Open Someday?
DeepSeek and Qwen publish open weights; MiMo-V2 begins life closed. MiMo-V2-Pro and Omni are, right now, proprietary API offerings, and the furthest Luo Fuli went was to say that some variant would be open-sourced down the road — no timeline attached. For the time being, that positions Xiaomi apart from the very Chinese labs it recruited from. The decision lines up with the split that was already visible at the start of 2026 (see “2026 Opening Outlook”): model releases keep accelerating while the open and closed camps drift apart, with Meta reported to be weighing a closed pivot for Avocado even as DeepSeek and Alibaba continue shipping downloadable weights. For developers who lean on open weights, Xiaomi today is a “callable, not downloadable” option.
What Developers Should Take Away
One: the API price war has drawn in another heavyweight — Artificial Analysis puts input pricing close to zero, and a trillion-class model at that price is a genuine win for the economics of RAG and agent pipelines. Two: OpenRouter’s anonymous leaderboards now double as a tryout arena for unreleased models, and keeping an eye on them lets you spot trends ahead of media coverage. Three: if closing first and opening later hardens into the norm, the wait before open-source communities receive weights will only lengthen; carry that lag as an architectural risk when you choose a model dependency.