On March 18, 2026, Xiaomi used a launch event in Beijing to unveil three in-house models at once: the flagship MiMo-V2-Pro, the omni-modal MiMo-V2-Omni, and a text-to-speech model, MiMo-V2-TTS. The next day, CEO Lei Jun committed to spending at least $8.7 billion on AI over the following three years, according to Reuters. A company known for phones and home appliances has stepped into the trillion-parameter race.
What makes this launch interesting is not just the money. MiMo-V2-Pro has more than 1 trillion total parameters, 42 billion active ones, and a 1-million-token context window — specifications in the same weight class as frontier-lab flagships. Lei Jun presented the technical details himself at the Beijing event. The more theatrical part is how it arrived: before the announcement, the model ran on OpenRouter under the pseudonym “Hunter Alpha,” topped the daily usage charts for several days, and processed over a trillion tokens. Users speculated it was a new DeepSeek system until Xiaomi confirmed it was an early internal test build of MiMo-V2-Pro. The sibling MiMo-V2-Omni accepts image, video, audio, and text in a single model, and MiMo-V2-TTS rounds out the trio for speech generation.
One Launch, One $8.7B Pledge
Reuters reported Lei Jun’s March 19 commitment of at least $8.7 billion over three years. Set against Xiaomi’s existing research budget of roughly 30 billion yuan per year, this is a real shift of resources across models, compute, and applications simultaneously. The model team is led by Luo Fuli, a former DeepSeek researcher who joined Xiaomi in late 2025. Poach a DeepSeek lead, then fund the infrastructure — Xiaomi is running a playbook that has already been validated.
The Hunter Alpha Playbook
Anonymous test releases are becoming standard choreography for new models: list under a codename, absorb real traffic free of brand halo and public-opinion pressure, then reveal the identity once the numbers look good. MiMo-V2-Omni did the same, listed as “Healer Alpha.” The pattern has a second benefit for the lab: a week of anonymous chart-topping is proof of demand that no internal benchmark can match, and it cost Xiaomi nothing but inference. For developers the practical lesson is simple: those unexplained but strong performers sitting atop OpenRouter charts deserve serious attention — they may be anyone’s next flagship, and pricing during the anonymous window is often aggressively cheap.
Trillion Parameters, 42B Active
MiMo-V2-Pro is a mixture-of-experts design: over 1 trillion total parameters, 42B activated per inference, and a 1-million-token context. That configuration targets agentic workloads — long documents, multi-tool calls, cross-session tasks — and the official positioning says outright that it is built to be “the brain of AI agents.” Independent evaluator Artificial Analysis already ranks it among the leading models on its intelligence index, and free trials were extended to April 2 through partners like OpenClaw because demand outran capacity.
Open or Closed?
Unlike the open-weights path taken by DeepSeek and Qwen, MiMo-V2 starts closed: both MiMo-V2-Pro and Omni are proprietary API products, and Luo Fuli would only say that some variant will be open-sourced later, with no timeline. That puts Xiaomi on the opposite side of the fence from the Chinese labs it hired from, at least for now. The choice echoes the divergence that was already visible at the start of 2026 (see “2026 Opening Outlook”): as model cadence accelerates, the open and closed routes are pulling apart, with Meta reportedly weighing a closed turn for Avocado while DeepSeek and Alibaba keep shipping downloadable weights. For developers who depend on open weights, Xiaomi is for now a “callable, not downloadable” option.
Three Practical Effects for Developers
First, the API price war gains another heavyweight: Artificial Analysis shows input pricing near zero, and a cheap trillion-class model is a real cost win for RAG and agent pipelines. Second, OpenRouter’s anonymous charts have become a proving ground for unreleased models — watching them surfaces trends earlier than press coverage does. Third, if “closed first, open later” becomes the norm, the lag before open-source communities get weights will keep growing; treat that lag as an architectural risk when you pick a model dependency.
Sources
- Xiaomi to invest at least $8.7 billion in AI over next three years — Reuters
- Xiaomi MiMo — Wikipedia
- Xiaomi unveils latest AI model — LinkedIn News
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.