Published on Mar. 30, 2026

"I call it a quiet ambush, not because we planned it that way, but because the shift from the chat paradigm to the agentic paradigm happened so fast that even we could hardly believe it," Luo Fuli wrote on X last week.

Luo previously worked as a researcher at DeepSeek. She is now at Xiaomi, where she leads the development of MiMo, the company’s proprietary large model.

The ambush Luo describes consisted of three simultaneous moves in artificial intelligence within a single week. In rapid succession, separate teams at Xiaomi unveiled a trio of large language models and a new electric vehicle equipped with the company’s own AI-powered driver assistance system.

Then founder Lei Jun took to a stage and announced AI investments totalling 60 billion yuan (roughly US$8.3 billion) over the coming three years.

Xiaomi was now on its way to becoming a "leader in physical AI with proprietary AI, operating system and chip capabilities", analysts at Goldman Sachs led by Timothy Zhao wrote.

A brief note on terminology. Physical AI and embodied AI are used almost interchangeably in the industry. Physical AI is the broader term, covering all AI that interacts with the physical world, whether in robots, vehicles, drones or IoT devices. Embodied AI is narrower, emphasising that the intelligence inhabits a body capable of moving and acting autonomously in the world. A humanoid robot is the textbook example. A temperature sensor in a smart home is not.

Xiaomi is a conglomerate that started as a smartphone maker, expanded into home appliances and the Internet of Things, and has been building electric vehicles since 2024, with considerable success in China’s fiercely competitive car market.

Now Xiaomi is also developing its own foundation models under the MiMo brand. This gives the company something that neither pure software firms such as OpenAI nor pure car manufacturers possess: an ecosystem of hundreds of millions of physical devices into which proprietary AI models can be integrated, models that do not merely compute or run inference in the cloud but can steer, navigate and interact with humans in the real world.

Xiaomi had been preparing its AI models in secret. A mysterious prototype had surfaced and set off speculation, but the announcement of a trio of proprietary large AI models at Xiaomi’s annual technology conference on 19 March caught the vast majority of analysts completely off guard.

Xiaomi unveiled MiMo-V2-Pro, MiMo-V2-Omni and MiMo-V2-TTS. The flagship MiMo-V2-Pro has more than one billion parameters and a context window of up to one million tokens. It ranks eighth globally on the Artificial Analysis leaderboard, ahead of Grok by xAI. MiMo-V2-Pro has already processed more than 1.5 trillion tokens, Reuters reported.

The second part of this AI ambush was the second generation of the Xiaomi SU7 electric sedan, which Lei Jun presented on the evening of 19 March in Beijing. The new vehicle’s driver assistance system is built on a proprietary AI model that, according to its maker, breaks the limits of data-driven imitation learning.

The new XLA architecture possesses "situational understanding" and is capable of "probabilistic reasoning", enabling it to "independently derive strategies in previously unknown complex scenarios", the Chinese automotive portal Gasgoo reported (in Chinese).

This upgrade connects driver assistance with embodied robotics for the first time, Gasgoo wrote, citing company spokespeople. Through multimodal inputs and the integration of the MiMo Embodied Model, the system is better at interpreting complex traffic situations and better equipped for long-tail scenarios.

The result, the company claims, is both more human-like decision-making and a range of new functions designed to be useful in everyday life.

Xiaomi says its upgraded HAD driver assistance system can now independently locate vacant parking spaces in underground car parks at shopping centres, identifying those closest to the driver’s preferred lift. After a voice command, the driver is navigated there and the SU7 parks itself. More than 3,000 shopping centres are already connected.

Lei Jun said Xiaomi’s AI development had progressed far faster than originally planned. "Our R&D and capital investment in AI will exceed 16 billion yuan (roughly US$2.2 billion) this year," he said.

Then came the third bombshell. Lei Jun announced that his company would invest more than 60 billion yuan (roughly US$8.3 billion) in AI over the next three years. Xiaomi’s share price swung wildly as observers tried to make sense of how it all fitted together.

Xiaomi intends to weave its three business strands of home appliances, vehicles and AI models into an ecosystem that Lei Jun calls "Human-Car-Home". An AI agent can execute more than 50 system functions from the car’s cockpit, from composing messages and managing calendars to controlling IoT devices at home. The vehicle becomes the first endpoint for embodied intelligence. The connected home becomes the second.

No other technology company has a comparable, multi-layered technology stack today. Tesla builds cars and trains driving models, and is also developing humanoid robots, but it does not have a broad ecosystem of vacuum cleaners, air conditioners or smartphones. OpenAI and Anthropic build powerful models but no hardware of their own. Xiaomi is present at both feasts of embodied AI and at the wedding.

In the week Xiaomi’s AI initiative became public, Chinese AI models led their American rivals in global usage for the third consecutive week. Five of the nine most-used models on the platform OpenRouter were Chinese (see also this week’s Deep Dive).

"We are living in an entirely new era. Both individuals and companies must embrace AI," said Xiaomi founder Lei Jun.