New NVIDIA Cosmos 3 Open Models for Interactive World Simulation

NVIDIA has added four open-weight checkpoints to its Cosmos 3 world model family, according to a Hugging Face post dated November 5, 2026, written by members of the company's team. All are post-trained versions of Cosmos 3 Nano with 16 billion trainable parameters, released under the OpenMDW1.1 license, which allows commercial and non-commercial use under its terms.

Two target autonomous driving. Transfer-Auto creates offline synthetic data, producing synchronized video across up to 11 camera views along with optional LiDAR, while developers vary weather, lighting and time of day. Sim-Auto is a distilled streaming version that generates camera and LiDAR output in small chunks using only a few denoising steps, though closed-loop use still requires a separate simulation runtime and renderer.

The other two focus on simulation and robotics. Sim-Transfer turns a physics simulator's depth, segmentation, edge or blur controls plus a text prompt into photorealistic video, leaving physics to the simulator. Sim-Bimanual predicts video directly from actions for two-armed YAM and AgiBot robots, driven by a policy or joysticks.

NVIDIA also published recipes for adapting the models to custom data, built on a four-stage process ending in self-forcing, which trains on generated history so compounding errors are addressed. The company presents the release as a shared base for robotics and driving research, while acknowledging that it remains to be measured where these models faithfully support policy learning and evaluation.