Xiaomi releases the source code of its first VLA model for robots

Xiaomi Opens Source Code of Its First VLA Model for Robots

Xiaomi opens source code of its first VLA model and changes the rules of the game in robotics

Xiaomi has taken a decisive step in developing autonomous systems by open-sourcing Xiaomi-Robotics-0 — its first Vision-Language-Action model for robots. The announcement, made on February 12, means that developers around the world will gain access to models with 4.7 billion parameters that combine visual perception, natural language understanding, and real-time robot control. This is not just another open-source project — it is a signal that the Chinese technology giant is seriously betting on the development of multimodal AI for physical automation and is ready to share its achievements with the global developer community.

Until now, the field of VLA models has remained the territory of the select few. The most powerful solutions in this area were created by companies like Tesla with their object manipulation apparatus and Google DeepMind with their robotics experiments. These models learn to connect what robots see with natural language commands and specific real-time actions. The problem is that training such systems requires a massive volume of video data, powerful computational resources, and deep understanding of robot mechanics. Xiaomi's opening changes this dynamic. By releasing the source code and weights of its model, the company democratizes access to a technology that was previously the prerogative of large corporations with enormous budgets.

Xiaomi-Robotics-0 is designed to work efficiently and practically. A model with 4.7 billion parameters is the golden mean between speed and performance. It is compact enough to run on robot platforms with limited computational capabilities, yet has enough power to solve complex manipulation tasks. The architecture combines three key components: a visual encoder that analyzes images from the robot's cameras; a language component that processes user instructions in natural language; and an action module that generates real-time control commands. This integration is critical because a robot must not only understand the task (for example, "pick up the red cube"), but also immediately translate it into movement coordinates and grip force.

Xiaomi's strategy of opening the code has clear logic. In the robotics ecosystem, competition unfolds not so much at the level of models, but at the level of hardware, software, and the application ecosystem. By opening the VLA model, the company allows thousands of developers to experiment with new use cases and adapt the technology to their own robots. This creates a network value effect: the more people improve the model and find new applications, the greater the likelihood that solutions emerging from this community will work with Xiaomi's in-house developments.

For the industry, this means accelerated development. Startups, universities, and corporations will be able to experiment with multimodal robot control without starting from scratch. The model is already trained on real data and has a reasonable architecture that can be adapted for different platforms. This is especially important at a moment when robotics stands on the threshold of transitioning from laboratories into production and everyday life.

Xiaomi's move demonstrates a broader trend: leaders in AI increasingly understand that opening some of their developments creates a stronger ecosystem than strict control. This does not mean the company is abandoning its robots — on the contrary, they will remain closed and proprietary. But the VLA model becomes the foundation on which an entire industry will grow, and Xiaomi is already laying the bricks in its foundation today.