What was launched and when
NVIDIA has confirmed the public availability of five specific models from the Cosmos3-Nano family. The publication date established in the official model cards is October 20, 2026. These models are not a single tool, but a set of specialized checkpoints that share a 16-billion-parameter architectural base.
The launch includes variants for driving simulation, visual control transfer, bimanual manipulation, and surgical robotics. All weights are hosted in NVIDIA's official repository on Hugging Face, with supporting code available on GitHub. This immediate availability allows research and development teams to begin integration without waiting for a commercial API.
Variants, modalities, and access
Each variant is designed for a specific use case. Cosmos3-Nano-Transfer-Auto and Cosmos3-Nano-Sim-Auto focus on driving scene generation, differing by their bidirectional or causal architecture, respectively. Cosmos3-Nano-Sim-Transfer converts simulation controls into RGB video, while Cosmos3-Nano-Sim-Bimanual simulates manipulation with two robotic arms.
The Cosmos3-H-Embodied-Nano variant stands out for its focus on surgical robotics, offering three operating modes: direct dynamics, inverse dynamics, and conditioned policy. Access is via model weight download, meaning the user must manage their own compute infrastructure. There is no published API fee; the cost falls on the GPU, memory, and system operation.
Summary of available variants
Comparative table of the five variants launched by NVIDIA.
| Model | Primary Focus | Inputs | Outputs |
|---|---|---|---|
| Cosmos3-Nano-Transfer-Auto | Bidirectional Driving | Text, Image, LiDAR | Video, LiDAR |
| Cosmos3-Nano-Sim-Auto | Causal Driving | Text, Image, LiDAR | Video, LiDAR |
| Cosmos3-Nano-Sim-Transfer | Control Transfer | Text, Video | Video |
| Cosmos3-Nano-Sim-Bimanual | Bimanual Manipulation | Text, Image, Actions | Video |
| Cosmos3-H-Embodied-Nano | Surgical Robotics | Text, Image, Actions | Video, Actions |
What changes for the user
For engineering teams, the main implication is the need for their own infrastructure. As these are 16B parameter checkpoints, inference requires high-performance hardware. Users cannot simply call an endpoint; they must deploy the model, manage video encoding, and ensure the stability of the generated sequence.
The distinction between causal and bidirectional models is crucial. Sim-Auto, being causal, allows for closed-loop simulation where the model responds to controls in real time, something Transfer-Auto does not offer in the same way. In robotics, the H-Embodied-Nano variant allows not only predicting video but also inferring the necessary actions, making it useful for training control policies.
Pricing, limits, and what remains to be verified
There is no published usage price, as the models are open-source under specific licenses (NVIDIA Open Model License or OpenMDW 1.1). The real cost is the infrastructure: it is estimated that next-generation GPUs are needed to handle the 16B parameters and video generation at 480p or 720p.
It is important to note that the published metrics (FID, FVD, PSNR, MSE) come from NVIDIA and have not been independently reproduced. Furthermore, photorealistic appearance does not guarantee physical consistency or safety. Users must validate temporal stability and control accuracy before using this data in critical systems.
Technical implications and licensing
The Cosmos3-Nano family relies on NVIDIA's cosmos-framework. The license varies by model: most use OpenMDW 1.1, while Cosmos3-H-Embodied-Nano is under the NVIDIA Open Model License Agreement. This implies specific restrictions for commercial use and code modification.
Integration requires attention to software dependencies and input control configuration. For example, in the driving variant, synchronization between multiview video and LiDAR is a technical challenge that the user must resolve. The availability of these models marks a step toward democratizing high-performance simulation, but demands considerable technical maturity from users.
Steps to deploy a Cosmos3-Nano model
Basic process to start using one of the published models.
- 01Clone the cosmos-framework repository from GitHub.
- 02Download the specific model weights from Hugging Face.
- 03Configure the runtime environment with GPU and CUDA dependencies.
- 04Define input parameters (prompts, images, controls).
- 05Run inference and validate the visual and data output.
Open questions
- The existence of a commercial API or token fees for these models has not been verified.
- Independent reproducibility of performance metrics (FID, PSNR, etc.) has not been performed by Inferama.
- Real performance on physical hardware or in out-of-distribution scenarios is not guaranteed by the official model cards.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction