NVIDIA Cosmos is an open platform for physical AI, providing world models, datasets, and tools for developers to build applications such as robotics, autonomous driving, and intelligent infrastructure. According to the official description, it aims to provide physical AI developers with better data, better environments, better starting points, and better tools.
Purpose and Positioning Cosmos focuses on Physical AI scenarios, and its target users are developers in fields such as robotics, autonomous vehicles, and intelligent infrastructure. The current model family, Cosmos 3, is a set of omnimodal world models based on a unified Mixture-of-Transformers architecture, which the official description summarizes as "one model family that can see, reason, simulate, and act."
Main Capabilities and Model Family Cosmos 3 provides three foundation models covering different deployment tiers, all sharing the same omnimodal architecture. The repository also includes example checkpoints in the cosmos3-examples collection. These are variants of the foundation models after post-training, used to demonstrate what specializations can be achieved through post-training with the Cosmos Framework. The official description states that they are capability demonstrations and are not part of the product line. In addition, the official release timeline shows: Cosmos3-Edge was released in July 2026, a 4B tier for real-time on-device deployment (Jetson AGX Orin / Thor / RTX Pro 6000); Cosmos 3 was released in May 2026, along with a Hugging Face collection and technical report.
Repository Responsibilities and Toolchain This GitHub repository is where the models reside, covering everything needed to explore, run, and evaluate Cosmos. Model training (SFT, LoRA, RL, distillation, etc.) is performed in the Cosmos Framework; data curation uses Cosmos Curator; and model output evaluation uses Cosmos Evaluator. Complete I/O specifications, generation settings, and supported action embodiments are available in the model reference documentation.
Usage and Runtime Environment The first run downloads the 16B checkpoint. Diffusion steps are computationally intensive, so long step times are normal. Before running the code, you must first request access to nvidia/Cosmos-1.0-Guardrail and accept its access conditions; after authorization, use the same account's Sign in with a Hugging Face read token; signing in alone does not grant access. Cosmos 3 can run on Diffusers, Transformers, vLLM, vLLM-Omni, SGLang, TensorRT-LLM, and NIM; see the environment setup guide for the specific backend choice. For image and audio modes and the full set of options, refer to the audio-video cookbook.
Use cases and limitations Suitable for exploring, running, and evaluating physical AI world models, including tasks such as video generation, aimed at areas such as robotics, autonomous driving, and intelligent infrastructure. Regarding limitations, the official documentation clearly states that Cosmos 3 may produce artifacts in long-duration, high-resolution, or physically complex outputs, such as temporal inconsistency, object deformation, and implausible dynamics; safety-critical applications require additional validation and system-level safety analysis; guardrails are enabled by default for generation.
Community and license Questions and ideas can go to Discussions, bug reports can be submitted to Issues, and code contributions are described in CONTRIBUTING.md. The source code and models are released under OpenMDW-1.1, and for custom licensing you can contact cosmos-license@nvidia.com. The project may download third-party open-source software, and you should review its license before use.
