“An open-source 3 billion parameter vision-language model used as a backbone for robotics training.”