Welcome, Spirit v1.5!
Spirit v1.5 is an open source robot foundation model developed by Spirit AI, a Chinese embodied AI startup also known as Qianxun Intelligence. Released on January 12, 2026, it belongs to the vision language action category of models for robotics. It is software intended for integration into robotic systems, not a physical robot. Its Humanoid Guide Brain Score of 5 out of 10 places it in the developing category. Shortly before its release, Spirit v1.5 ranked first overall on RoboChallenge, a benchmark that evaluates models using real robots rather than simulation alone. That result provides a specific evaluation milestone, while the broader score reflects a model still in development. A central feature of the project is its training approach, which emphasizes varied data collected through less scripted operator activity. Spirit AI released the model and code openly, positioning Spirit v1.5 as a foundation that other robotics teams can examine and build upon.
The training methodology centers on how demonstrations are collected. Instead of restricting operators to clean, tightly scripted sequences, Spirit AI allowed them to pursue open goals using their own approaches. This process generated messier and more varied examples, making data diversity a defining element of the project. The company argues that such variation supports faster learning of new tasks, an intended benefit rather than a guarantee for every deployment. RoboChallenge adds a separate evaluation context by testing models on physical robotic hardware. Its importance lies in the distinction between benchmark results obtained through interaction with real equipment and results obtained entirely within simulated environments.
Spirit AI makes Spirit v1.5 and its code openly available so that other teams can build on the release. The project forms part of the company’s broader ambition to develop a universal software foundation for home robots, including its own Moz humanoid. That ambition describes the direction of the program, not a completed universal solution. The January 2026 release and its RoboChallenge ranking make Spirit v1.5 a notable entry in the developing robot foundation model landscape. Its broader significance rests on the combination of open distribution, an emphasis on varied operator data, and evaluation using real robots. Together, these elements establish a concrete starting point for further development and independent investigation.
