NVIDIA posts four Cosmos 3 Nano checkpoints for interactive world simulation

NVIDIA’s Cosmos world-model platform added four post-trained Cosmos 3 Nano checkpoints on 5 November 2026, published on Hugging Face. Each has 16B trainable parameters. The models use OpenMDW1.1, which permits commercial and non-commercial use under its terms; do not shorthand that as “fully open.”

The four checkpoints

• Cosmos3-Nano-Transfer-Auto. Offline synthetic driving data: synchronized video across up to 11 camera views (including fisheyes), optional joint RGB–LiDAR, with World Scenario Map controls for scene structure and traffic motion, and text prompts for appearance.
• Cosmos3-Nano-Sim-Auto. The causally distilled streaming variant of Transfer-Auto. It generates successive chunks of synchronized camera video and LiDAR in a few denoising steps for interactive rollouts.
• Cosmos3-Nano-Sim-Transfer. Turns a physics simulator’s spatial controls (depth, segmentation, edges, blur) plus a text prompt into photorealistic RGB the policy can consume, while physics stays with the simulator.
• Cosmos3-Nano-Sim-Bimanual. Action-to-video for two-arm YAM and AgiBot embodiments, driven by a policy or joysticks. One control sequence can switch between viewpoint changes and manipulation.

NVIDIA also posts post-training recipes from Cosmos3-Nano (or compatible base weights) through bidirectional adaptation, teacher forcing, consistency distillation and self-forcing, plus agent-runnable skills and notebooks.

What you need to run them

Requirements listed in the post: Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit. The published vLLM-Omni setup was tested on one GB200; that is not a minimum-memory specification.

Why this matters

If you are wiring NVIDIA stacks into cloud robotics pipelines, as in the AWS physical-AI toolchain sample, these checkpoints give downloadable transfer and streaming sim pieces rather than only a managed recipe. Against closed driving world models such as Baidu’s Apollo World, Cosmos 3 Nano is the open-weight route with explicit sensor and control contracts. Interactive media world models such as Genie 4 chase different surfaces (audio, live text steering); here the beat is AV multi-view and bimanual robot observation. Pair any policy you train with real-robot checks such as RoboColosseum before you trust a sim score.