NVIDIA releases four open Cosmos 3 simulation models

NVIDIA published four new open-weight checkpoints for its Cosmos 3 model family, according to a post on Hugging Face dated November 5, 2026. All four are post-trained versions of Cosmos 3 Nano and carry 16B trainable parameters.

What was released

• Cosmos3-Nano-Transfer-Auto produces synthetic driving data offline. It can generate synchronized video across up to 11 camera views, fisheye cameras included, and can add LiDAR. Text prompts steer weather, lighting and time of day.
• Cosmos3-Nano-Sim-Auto is the streaming version of that model. It produces short chunks of camera and LiDAR output in a few denoising steps, so a simulation can run long and respond to changing controls.
• Cosmos3-Nano-Sim-Transfer takes spatial controls from a physics simulator, such as depth, segmentation and edges, plus a text prompt, and renders color video a policy can consume. Physics and ground-truth state stay in the simulator.
• Cosmos3-Nano-Sim-Bimanual predicts the next video frames from robot or camera actions. It supports two-arm setups for the YAM and AgiBot embodiments and can be driven by a policy or by joysticks.

Licensing and requirements

The weights use the OpenMDW1.1 license, which permits commercial and non-commercial use under its terms. Running the Bimanual checkpoint requires Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit. NVIDIA says the published setup was tested on one GB200 and that this is not a minimum-memory figure.

Adapting the models

NVIDIA also publishes post-training recipes that cover four stages, from adapting to a domain through distillation to training on the model's own generated history. Teams can adapt Transfer-Auto to their own camera and LiDAR layout, then distill it into a streaming model. The post advises evaluating on held-out episodes, because a short successful run proves the pipeline executes without showing the model has converged.

The post is candid that the hard problems remain open: the next step is to measure where these models support policy learning and evaluation and where they fall short.

What it means

This is a research and engineering release for robotics and autonomous-driving developers. It does not touch the sales, support or operations workflows most B2B teams automate.