AI large models are evolving from a product into the infrastructure a company runs on. Xiaomi recently reorganised XiaoAI and its AI business: the near-decade-old assistant is shifting from a standalone team to one backed by the base-model team, a cloud engineering team and an on-device OS team. Reporting says the MiMo base-model team led by Luo Fuli supplies the underlying model, Luan Jian builds the cloud engineering that runs MiMo on XiaoAI, and on-device capability moves to the phone OS team. The former XiaoAI lead Wang Gang has moved on.
MiMo’s priority rises again
Before MiMo, the AI lab, XiaoAI, phone OS, auto and IoT teams each held capable AI. MiMo did not instantly become the uniform choice, and teams may still pick models by effect, cost and ship pace. But as a long-term strategy MiMo must reach phones, cars, homes, wearables and robots, becoming an underlying capability every line can call. Xiaomi’s businesses sit across many terminals; if each builds its own AI, experience fragments and the “human-vehicle-home” ecosystem collapses. This reorganisation gives MiMo a shared understanding base that each system schedules for its scene.
Xiaomi’s strong internet culture, frequent cross-team talk and fast resource assembly, is what lets MiMo close its loop: the model team iterates only if business teams feed back scene data, user feedback and real call chains. That binding is a cost, the price of a self-built base model, but without it MiMo becomes an expensive project drifting from the business.
A key step to hold the ecosystem entrance
Building a base model looks costly with uncertain return, yet in the AI-agent reshaping of entrances, services and value distribution it may be Xiaomi’s most far-sighted bet. The phone industry consensus says makers need not build base models, just do on-device work and plug in external models. The logic is sound: training, inference, talent and compute are long-term costs a hardware firm’s cash flow struggles to fund at internet intensity.
But for over a decade phone makers sold hardware while profit also came from preloads, app stores, browsers, ads, search, game distribution and finance. Agents rewrite that: users will increasingly just tell an agent the task, compressing the old entrance chain. A source tells LeiFeng Net agent-to-agent collaboration will come, Xiaomi’s agent handing part of a request to Tencent, Alibaba, Meituan, Ctrip or Douyin agents, yet Tencent will not let MiMo call WeChat’s core interface, and Xiaomi will not let outside agents control its hardware.
Whoever holds the user entrance, key data and fulfilment keeps the louder voice. A hardware maker without its own agent becomes mere hardware under someone else’s intelligent entrance, with outside models deciding what service to show, which platform to call, how to rank, how to charge. So Xiaomi needs an intelligent entrance representing its ecosystem. Even if MiMo is not the biggest or highest-ranked model, as long as it plays the agent inside Xiaomi’s ecosystem it defends the firm’s future service distribution. Short term MiMo burns cash; long term, if agents become the next entrance, not building it could cost Xiaomi the entrance itself.
Editor’s note: This is an adapted translation of the original LeiFeng Net report. It has been trimmed and restructured for readability for an international business audience.