Cover: AI-generated editorial illustration.
On November 5, 2026, NVIDIA published four new open-weight checkpoints under the Cosmos 3 Nano family, each targeting a specific simulation gap in autonomous driving or robotics research.
The four models:
- Cosmos3-Nano-Transfer-Auto: Offline synthetic driving data generation across up to 11 camera views including fisheyes, with joint RGB and LiDAR output.
- Cosmos3-Nano-Sim-Auto: A causally distilled streaming variant of Transfer-Auto, producing successive chunks of synchronized camera video and LiDAR for interactive closed-loop rollouts.
- Cosmos3-Nano-Sim-Transfer: Takes spatial controls from a physics simulator (depth, segmentation, edges) plus a text prompt and renders color observations a policy can consume directly.
- Cosmos3-Nano-Sim-Bimanual: Predicts the next visual observation from bimanual robot actions, supporting YAM and AgiBot embodiments via policy or joystick control.
All four have 16 billion trainable parameters and are released under OpenMDW 1.1, which permits commercial use under its terms. NVIDIA also publishes the four-stage post-training recipe so teams can adapt any checkpoint to their own sensor rigs or robot embodiments.
The practical distinction between Sim-Transfer and Sim-Bimanual matters for how you structure your stack. Sim-Transfer offloads physics to an existing simulator and renders visuals on top; Sim-Bimanual predicts observations entirely from actions without a separate physics engine. Teams with an existing sim already in place will likely reach for Sim-Transfer first; those building from scratch may find Sim-Bimanual a faster starting point.
Requirements include Linux, an NVIDIA GPU, Docker, and the NVIDIA Container Toolkit. NVIDIA tested on a GB200 but does not state that as a minimum. Weights are available via Hugging Face.