Cosmos Framework is NVIDIA’s open, end-to-end framework for training and serving world models, including the Cosmos 3 model family. World models learn how physical scenes evolve, which makes them central to robotics and physical AI development. The project has about 568 stars on GitHub and was actively updated in October 2026.
Everything lives in a single top-level Python package, cosmos_framework. On the training side it provides a distributed trainer combining FSDP with tensor, context and pipeline parallelism, native distributed checkpoints with Hugging Face safetensors import and export, and dataset adapters for JSONL, WebDataset and LeRobot formats, the last of which is widely used for robot demonstration data. On the inference side it supports Diffusers, Transformers and vLLM backends, with offline batch generation and online serving. The framework is part of the wider NVIDIA Cosmos project family and is the training and serving companion to the Cosmos model releases on Hugging Face, aimed at teams building simulation and world-model pipelines for physical AI.
NVIDIA positions Cosmos as its platform for physical AI, and this framework is where the training and serving code for those models lives, so robotics teams can fine-tune world models on their own scene and robot data instead of using the base models exactly as released.
Cosmos Framework is publicly available on GitHub from NVIDIA at no cost; review the repository for its license terms, and expect to supply your own GPU infrastructure for training.
Cosmos Framework is for robotics and physical-AI teams that train or fine-tune world models and need a complete, distributed training and serving stack from the model’s creators.