Xiaomi XR-1 is an open-weights vision-language-action robotics model for research labs and hardware startups building embodied agents. Xiaomi Robotics published the 34B-parameter model with weights on Hugging Face and code on GitHub, trained on more than 100,000 hours of real-world manipulation trajectories rather than simulation alone.
Core Features
- A 34B-parameter vision-language-action design that maps camera input plus a text instruction directly to robot actions.
- Pairs a pre-trained Qwen3-VL vision-language backbone with a Diffusion Transformer action head through a Mixture-of-Transformers architecture.
- Two-stage pre-training and alignment borrowed from LLM practice, instead of task-by-task behaviour cloning.
- Demonstrated out-of-the-box mobile manipulation in environments the model never saw during training.
- Open weights on Hugging Face with a public GitHub repository, so the policy can be fine-tuned on your own trajectories.
Use Cases / Best For
- Robotics labs that want a pre-trained VLA baseline rather than training manipulation policies from scratch.
- Startups adapting a general manipulation policy onto a specific arm or mobile base.
- Researchers testing whether LLM-style pre-training scaling laws carry over to physical task completion rates.
Pricing
Free. The weights are published openly on Hugging Face with no subscription and no per-token API fee. Your real cost is hardware: a 34B backbone plus a diffusion action head needs serious GPU memory to fine-tune, and you still have to supply robots and collect trajectories.
Our Take
Best for teams that already own hardware and want a strong starting checkpoint, not for anyone hoping to drop a model onto a robot over a weekend. The interesting claim is that pre-training scaling laws transfer to real-world completion rates, and Xiaomi published enough detail to test it. The trade-off against closed policies from Figure or Physical Intelligence is support: you get the weights and nothing else. Browse more open releases in AI models and engines, or compare with GLM-5.2.
