Bagel Labs is an embodied-AI research lab focused on physical AI and world models, building general world-action models for robot learning and control. It uses compact model architectures and efficient training to speed up iteration across model generations, releases WorldDiT, a robot world-action model without a VLM backbone, and open-sources models and tools such as the PARIS and PARIS 2.0 diffusion models and ZKLoRA zero-knowledge training verification on HuggingFace and GitHub.