AI large models are evolving from a business and product into the infrastructure a company runs on.
Past software and hardware firms organised by product, with AI as one function among imaging, voice, recommendation or interaction. After foundation models arrived, AI reached product definition, user interaction and service distribution. Teams once split by product must re-coordinate around model capability.
Globally, firms made similar moves. Google merged DeepMind and Brain. Meta raised AI’s internal priority toward agents. In China, Tencent folded more AI into CSIG. ByteDance split Seed and Flow for model and application.
Xiaomi recently reorganised XiaoAI and AI-related work. The near-decade-old product is shifting from a relatively independent team to support from a base-model team, a cloud engineering team and an on-device OS team.
Domestic media reports that after the reshuffle, XiaoAI’s tech stack splits three ways: the MiMo base-model team led by Luo Fuli provides the underlying model. Luan Jian leads MiMo’s external engineering on XiaoAI. On-device capability once owned by XiaoAI moves to phone and other sub-business OS teams. Wang Gang, former XiaoAI lead, has reportedly moved on. Luan Jian, a key AI figure, joined Xiaomi’s large-model team in April 2023 and turned to Agentic AI after Luo Fuli took over end of 2025.
MiMo priority rises again
Before MiMo, AI labs, XiaoAI, phone OS, automotive, IoT and XiaoAI teams had each built AI capability in their scenes. After MiMo, it did not instantly become every team’s uniform choice. Teams may still pick models by effect, cost and schedule. But for Xiaomi, MiMo cannot be merely optional. As a major long-term strategy it must reach phones, cars, home, wearables and robots, becoming a base layer every line can call.
Xiaomi’s business spreads across many terminals and systems. If each line builds its own AI, it ends with a car assistant, a phone assistant, a home assistant with different logic. After the experience fractures, the human-car-home ecosystem is impossible. This reshuffle fixes that: different terminals share a similar understanding base from MiMo, then phone, car and IoT systems schedule by scene.
Xiaomi’s internet-style organisation, with frequent cross-team communication, lets it quickly form a strike force toward a clear goal, its old strength: fast resource concentration after direction is set, then product definition, supply chain, channel and execution to arrive late yet lead.
This binding also pressures. MiMo must satisfy phones, cars and IoT on stability, response, cost and security, and business teams must accept deeper coordination. For Xiaomi this is extra organisational cost, the price of a self-built base model. Only when the loop forms does continued base-model investment make sense. Otherwise MiMo becomes a costly project drifting from business.
The key step to hold the ecosystem entry
Continuing to spend on a base model looks costly with uncertain return. But in the rethink of AI-Agent entry, service and profit distribution, this may be Xiaomi’s most strategically far-sighted bet of the AI era.
Inside phones there is a consensus that makers need not build base models, that the realistic path is strong on-device capability, system features and scenes, then plug in external models via hardware and OS ecosystems. The reasoning is solid: base models cost too much in training, inference, talent and compute. Makers are hardware firms whose cash flow hardly sustains an internet-level spending war.
But for over a decade, makers sold hardware while profit came not only from hardware margin. Pre-installed software, app stores, browser entry, ads, search, game distribution and finance referral are all profit. AI Agent disrupts this by changing how users reach services.
Previously, to shop, hail a ride or book, users opened an app, browser or store. Makers joined value distribution through entry, distribution, recommendation and ads. In future, as users task Agents directly, the original entry chain compresses. Users need not open the store, pass the browser, or click ads in feeds. The Agent calls services by intent, finishes the task, returns the result.
A knowledgeable source says Agent-to-Agent collaboration will surely appear: different firms’ models connect, exchange intent, pass results, complete cross-platform tasks. Xiaomi’s Agent could understand a need and hand part to Tencent, Alibaba, Meituan, Ctrip or Douyin Agents. But Tencent will not let MiMo call WeChat’s core interface, nor Xiaomi allow external Agents to control its hardware. Collaboration is still inter-firm profit distribution. Whoever holds the user entry, key data and service fulfilment holds more say.
Without its own Agent, a hardware maker becomes only the hardware carrier under another’s intelligence, with external models deciding what services to show users, which platform to call, how to rank, how to charge. For Xiaomi, if the model service becomes a new intermediary between user and device, its years-built human-car-home ecosystem could be redistributed at the intelligence layer.
So Xiaomi needs an intelligent entry representing its ecosystem interest. Even if MiMo is not the biggest or highest-charting model, as long as it plays the Agent role in Xiaomi’s ecosystem, it helps hold the future service-distribution position.
The early Xiaomi-Huami relationship is a clean example. Huami supported Mi bands and watches, but both sides long tugged over health-data ownership. The same issue will recur more often across AI, cloud, IoT and automotive. Data ownership becomes a new profit-distribution question, more complex than before.
Short term, MiMo burns cash with no clear revenue. Long term, if AI Agent truly becomes the next entry, not building a base model today may cost more in lost entry tomorrow.
Images
Editor’s note: This is an adapted translation of the original LeiFengWang report. It has been trimmed and restructured for readability for an international business audience.