The repository offers a comprehensive toolkit for working with AI models throughout the entire lifecycle. It includes refactored training and inference code for diffusion models like Cosmos-Predict2.5 and Cosmos-Transfer2.5, PyTorch reproductions of Pi0/ Pi0.5 for vision-language-action tasks, and infrastructure for VLA foundation models. The platform supports model compression and deployment alongside training workflows, serving as a unified hub for researchers and developers working with generative and perceptual AI systems.