A phone maker that also builds cars has started giving away the brains it designed for robots. Xiaomi's recent releases suggest the company is treating embodied intelligence (具身智能) as a core platform rather than a side experiment, and it is doing so in the open.
The move puts a consumer-hardware giant directly into a field long dominated by labs and startups: teaching machines to act, not just to answer. For a company whose business is shipping millions of devices, opening the model weights is a calculated bet that the ecosystem, not the model, is where value will accumulate.
Two models, one robotics stack
In July 2026, Xiaomi launched and open-sourced Xiaomi-Robotics-U0, a 38-billion-parameter multimodal autoregressive generative foundation model built for embodied intelligence. The company says it ranked first among 126 models on the WorldArena benchmark.
Days later it released Xiaomi-Robotics-1, an embodied foundation model aimed at real-world mobile manipulation. This model was pre-trained on 100,000 hours of real-world data and, Xiaomi reports, ranked first on the RoboCasa365 simulation benchmark.
Together the two cover different ends of the problem. U0 is closer to a "world model" that can generate and reason about scenes; Robotics-1 is tuned for the messy reality of moving and handling objects. Releasing both under open licenses signals that Xiaomi wants outside developers, universities, and hardware partners to build on top of its stack.
What the models actually do
The U0 model is described as a unified generative system that handles four task families: embodied scene generation, embodied transfer, embodied video generation, and general image-text generation. The idea is to give robots a richer internal representation of space and action, not just a language interface.
Robotics-1 takes a different route. Rather than depending on scarce real-robot demonstrations, it uses a two-stage recipe: large-scale pre-training on off-body trajectory data, then fine-tuning on a cross-embodiment dataset that includes real robot hours. Xiaomi says the gains from pre-training scale cleanly into the real-robot tasks, which matters because high-quality robot data is the field's chronic bottleneck.
Neither model is a finished product. They are foundations — weights and code that lower the entry cost for anyone trying to make a machine operate in the physical world.
MiMo's moment on the global leaderboard
The robotics releases sit on top of Xiaomi's language-model line. MiMo-V2.5, the foundation model launched in April 2026, drew widespread global usage, and the company's interim report states it ranked first in weekly token usage on the OpenRouter platform, reaching 10.5 trillion tokens.
That figure matters less as a trophy than as evidence of distribution. A model that sits inside phones, cars, and home devices inherits a usage base no pure-play lab can match. In June 2026, Xiaomi also unveiled Miloco 2.0 (Xiaomi Local Copilot 2.0), an open-source home-intelligence solution powered by MiMo that aims at "proactive" smart homes that remember, recognize people, and take action.
The throughline is consistency: one model family feeding screens, speakers, cars, and now robot bodies.
Robots already on the factory floor
Xiaomi is not only publishing models; it is running them. At its electric-vehicle factory, the company demonstrated robots at a self-tapping threaded insert nut-loading station, where dual-side task success rose from 90.2 percent to 98 percent. The robots were also placed at new logistics workstations in the final-assembly workshop — center-console side-cover sorting and returnable-box folding — each reported at a 90 percent success rate.
In June 2026, Xiaomi's robotics team took first place in both the GigaBrain Challenge-RoboChallenge track at the CVPR 2026 Workshops and the Whole Body Control track at ICRA 2026, which the company cites as evidence of frontier-level research, not just marketing demos.
The R&D bill behind the open release
None of this is cheap. Xiaomi's 2026 interim report shows research and development spending rose 25.6 percent year over year to 18.2 billion yuan (about US$2.56 billion, HK$20.02 billion) in the first half. Research staff made up 47.2 percent of total employees, and the company held more than 47,000 patents worldwide.
That scale is the quiet headline. Open-sourcing models is affordable precisely because the parent can absorb the training cost across a hardware empire. In August 2026, Xiaomi also shipped a beta of HyperOS 4 with an upgraded "Hyper XiaoAi 2.0 (超级小爱2.0)" assistant that ties the MiMo model into cross-device task orchestration — another thread linking the model to shipped products.
Trend reading
According to Li An, chief scientist at BrainNet (脑机网), open-sourcing robotics models is less about generosity and more about building a developer ecosystem that will define the next interface between software and physical machines, where the winner owns the standard, not just the demo.
Honest limitations
The benchmark rankings (WorldArena, RoboCasa365) are self-reported in the interim report and reflect specific evaluation protocols, not guaranteed real-world performance across every task a robot might face. Factory success-rate figures come from Xiaomi's own demonstrations at its EV plant and have not been independently audited. Open-source release of model weights is not the same as turnkey deployment; integration, safety certification, and hardware cost remain real barriers for most adopters.
What readers can do now
- Robotics startups should evaluate Xiaomi-Robotics-U0 and Robotics-1 weights as a free baseline before committing to custom embodied models, and benchmark them against their own tasks.
- Hardware makers can read the open releases as a signal that the "model plus body" integration layer is becoming a commodity, and differentiate on reliability, certification, and total cost instead.
- Enterprises weighing in-house AI should note Xiaomi's reported R&D scale (18.2 billion yuan, about US$2.56 billion) as a realistic benchmark for what sustained model investment demands.