Xiaomi open-sources its robotics world model with training code
Xiaomi-Robotics-U0 is now open together with its training framework, under Apache 2.0. Video generation is listed among its capabilities, but the model card says the video checkpoint has not shipped.
Xiaomi has released the training and inference framework for Xiaomi-Robotics-U0, its world model for robotics, along with weights at 4B and 34B. The licence is Apache 2.0, so commercial use, modification and redistribution are all allowed. The code is on GitHub and the weights are on Hugging Face.
The announcement lists the supported tasks: embodied scene generation and editing, single-image generation and image editing, video generation, and interleaved vision-language sequence generation.
What is actually new today is the training side
The news section of the model card states that the technical report, the weights and the inference code all shipped in July 2026. The weights have been out for two months. What today's announcement adds is the training half: there is now a documented path to train the model on your own data.
That is not a small distinction. The promise of a world model is not to produce finished output but to multiply data for robots, and without a training framework that promise stayed half-delivered.
Video generation is on the list, but the file is missing
Video generation appears as a bullet in the announcement. The model card sets a clear limit on it: the required Xiaomi-Robotics-U0-Video checkpoint is "coming soon". Anyone who clones the repository today cannot do the video task the list advertises.
And the video in question is not video in the cinematic sense. In the model card's own words it is an "embodied video rollout": starting from an initial observation and predicting how the scene around the robot will unfold. The point is not a good-looking shot; it is generating data to train a robot on.
The one speed figure on offer
The repository publishes a single number, and it belongs to the image side: on one H20, FlashAR with vLLM reaches 5.44 s/img for 1024x1024 text-to-image. The same measurement puts that at 82.86 times faster than plain autoregressive execution and 3.04 times faster than FlashAR eager.
No VRAM requirement is stated anywhere. It is also worth noting that on Hugging Face the model announced as "4B" is listed at 5 billion parameters, the large one at 34 billion, and the FlashAR variant at 38 billion.
Why it matters
In robotics the bottleneck is not the model but the data: collecting it with real robots in the real world is slow and expensive. That is why world models draw attention, they simulate the scene and multiply the data. A world model under Apache 2.0, with weights in your hands, moves that work out of the lab.
Provided every item on the list has something behind it in the box. In this announcement, one item does not yet.
