Skip to main content

Mingxi Intelligent Open Sources MiniCPM-Robot: A 1.5B VLA Model for Personal Developers

For a long time, embodied intelligence was a playground reserved for big companies and well-funded labs. The models were massive, the barriers high, and ordinary developers could only watch from the sidelines. But on July 19, OpenBMB kicked the door wide open. They released the first series of embodied AI models, MiniCPM-Robot, fully open-sourcing a 1.5B parameter vision-language-action (VLA) model. Now, individual developers can run operations and target tracking on real robots. This level of openness is rare in the field.

The series includes three core components. The main one is MiniCPM-RobotManip, a 1.5B parameter general VLA model designed for robot manipulation. Then there's MiniCPM-RobotTrack, a compact model for real-world target tracking. And to tie it all together, OpenBMB also released PhyAI, a high-performance inference framework built specifically for embodied models. The goal is straightforward: enable robots to truly understand, remember, and act. Together, these components bring efficient, practical, and open embodied intelligence from papers and demos into real-world robots.

Image

For developers, the best part is how accessible it is. Model weights and code are hosted on GitHub at github.com/OpenBMB/MiniCPM-Robot, and the weights for MiniCPM-RobotManip and MiniCPM-RobotTrack are also available on Hugging Face. Anyone working on robots can clone the repository and start running experiments immediately.

When a 1.5B VLA model is no longer locked in the lab, the pace of embodied intelligence may exceed many people's expectations. The project address is https://github.com/OpenBMB/MiniCPM-Robot.

Key Points

  • Open Source VLA Model: MiniCPM-Robot is a fully open-source 1.5B parameter vision-language-action model for embodied AI.
  • Three Components: Includes MiniCPM-RobotManip for manipulation, MiniCPM-RobotTrack for tracking, and PhyAI inference framework.
  • Accessible to Individuals: Model weights and code are freely available on GitHub and Hugging Face, enabling personal developers to run real robot operations.
  • Breaking Barriers: This release lowers the entry barrier for embodied intelligence, previously dominated by large companies and labs.