NeuroAI NEUROAINEUROAI.SITE
ESC

Xiaomi's CyberOne and MiMo show a phone maker betting AI belongs in bodies, not just chats

Xiaomi unveiled its CyberOne humanoid (人形机器人) back in 2022, but the real pivot came in 2025: it open-sourced MiMo, a reasoning large model (大模型), and then MiMo-Embodied, a 7-billion-parameter model that links robot control to self-driving. The thread is one idea — put the same brain in a robot, a car and a home.

2026-09-28 · 766 words · NeuroAI
Xiaomi's CyberOne and MiMo show a phone maker betting AI belongs in bodies, not just chats

Xiaomi made its name putting computers in people's pockets. In 2025 it started putting its models into things that move.

The company that sells phones and EVs now talks about "embodied intelligence (具身智能)" as a core strategy — and it is backing the words with open weights.

The humanoid that started it

Xiaomi first showed CyberOne (人形机器人) in 2022, weeks before Tesla showed its Optimus prototype. At debut it walked and handled objects more smoothly than Tesla's first attempt, though Tesla has since pulled far ahead. CyberOne later appeared at MWC 2023 in Barcelona.

CyberOne was a statement, not a product. The interesting part was the doctrine: Xiaomi wanted one intelligence that could live in a phone, a car and a robot.

MiMo: the model with a reasoning habit

On 30 April 2025, Xiaomi's Core LLM team open-sourced Xiaomi MiMo, the company's first reasoning large model (大模型). At just 7 billion parameters, MiMo beat OpenAI's o1-mini and Alibaba's Qwen QwQ-32B-Preview on math (AIME 24-25) and code (LiveCodeBench v5) benchmarks — small model, outsized score.

In December 2025 Xiaomi released MiMo-V2-Flash, an open-source MoE model with about 309 billion total parameters and 15 billion active, under the permissive MIT licence. Around the same time, Luo Fuli (罗福莉) — a former DeepSeek researcher — joined as MiMo team lead, a hire that signalled Xiaomi was serious about closing the gap to frontier labs.

MiMo-Embodied: one brain for robots and cars

The pivot point was 21 November 2025, when Xiaomi released and fully open-sourced MiMo-Embodied, a 7-billion-parameter vision-language model. What makes it unusual:

  • It is described as the industry's first cross-domain foundation model linking embodied intelligence (具身智能) and autonomous driving.
  • It handles both robot tasks — affordance reasoning, task planning, spatial understanding — and driving tasks — perception, state prediction, driving planning.
  • Xiaomi says it set new marks on 29 core benchmarks (17 embodied, 12 autonomous-driving).
  • Weights are on Hugging Face (XiaomiMiMo/MiMo-Embodied-7B).

The bet is integration. Instead of a robot brain and a car brain built separately, Xiaomi wants one model that understands a room and a road the same way — so a home robot and an SU7 learn from each other.

Why a phone company does this

Xiaomi's "human–car–home" (人车家) strategy is not a slogan. A unified embodied model is the glue: the same agent architecture that parks your car could, in theory, fold your laundry. Xiaomi has said it plans to deploy MiMo-Embodied across robot vacuums, factory vehicles and its SU7 — though concrete shipping dates for those uses have not been confirmed.

Opening the weights is the Android play: lower the barrier, attract developers, and hope an ecosystem forms around Xiaomi's hardware the way Android formed around phones.

The honest gaps

CyberOne has not become a product you can buy. MiMo-Embodied's benchmark leads are Xiaomi's own claims, not yet reproduced widely by independent labs. And "one brain for everything" is still a research thesis, not a shipped feature. Xiaomi is early — but it is early with real weights on GitHub.

What readers can do now

  • If you build robots or autos: pull MiMo-Embodied-7B from Hugging Face and test the cross-domain claim on your own data; a 7B model is cheap to run.
  • If you track Chinese AI: watch whether Xiaomi keeps MiMo open. Closed weights would break the ecosystem story.
  • If you use Xiaomi gear: expect MiMo to show up first in SU7 driving assistance and home devices — the visible edge of a silent rebuild.

Honest limitations

Core facts — CyberOne humanoid (人形机器人) unveiled 2022 (weeks before Tesla AI Day, 30 Sep 2022), shown at MWC 2023; Xiaomi MiMo open-sourced 30 April 2025 as a 7B reasoning large model (大模型) by Xiaomi's Core LLM team, beating o1-mini and Qwen QwQ-32B-Preview on AIME/LiveCodeBench per company claims; MiMo-V2-Flash released ~16 Dec 2025, ~309B total / 15B active params, MIT licence; MiMo-Embodied released and open-sourced 21 Nov 2025, 7B vision-language model, cross-domain embodied+autonomous-driving, 29 benchmarks (17+12), on Hugging Face — come from Beijing Municipal S&T Commission (kw.beijing.gov.cn, 1 Dec 2025), STAR Market Daily (科创板日报, 21 Nov 2025), Shanghai Securities News/cnstock (17 Dec 2025), and Baidu Baike. Luo Fuli (罗福莉), ex-DeepSeek, joined Xiaomi Nov 2025 as MiMo lead — reported by multiple Chinese outlets. The benchmark "leads" and the "first cross-domain model" claim are Xiaomi's own and not yet independently reproduced. CyberOne remains a demonstration, not a sold product. Planned deployments in vacuums/SU7/factory vehicles are company-stated intent without confirmed dates. No RMB funding amounts in this article; no currency conversion required. Not investment advice. Current to 28 September 2026.

Related coverage

More in “Companies & Stack” → · Back to home · Markdown version