NVIDIA's physical-AI flagship, billed as the "world's first fully open omnimodel": a unified Mixture-of-Transformers combining autoregressive and diffusion generation across text, image, video, audio, and action, trained on ~20T multimodal tokens. Ships in Super and Nano variants with a "Cosmos Coalition" of robotics and simulation partners. Launched at GTC Taipei June 1 (backfilled); the technical report was revised through late June. The companion Cosmos3-DROID release repackages the DROID robot corpus in LeRobotDataset v3.0 format — 71,907 episodes / 22M frames / 707GB under OpenMDW-1.1.

On October 20, 2026 NVIDIA released four Cosmos3-Nano post-trained checkpoints for interactive simulation (Cosmos3-Nano-Transfer-Auto, -Sim-Auto, -Sim-Transfer and -Sim-Bimanual, OpenMDW-1.1), and on November 4 a technical report on them, "Interactive World Simulation with Cosmos for Physical AI". The report adapts the model to stream observations in response to actions and scene controls, covering camera navigation, robot manipulation, controllable video transfer and multi-view driving. Its model card describes Cosmos3-Nano-Sim-Auto as a 16B-trainable-parameter Mixture-of-Transformers derived from Cosmos3-Nano in two stages (post-training on multi-sensor driving data, then causal distillation to a few-step chunked generator) that renders up to 11 camera views with aligned LiDAR. These are application checkpoints of Cosmos 3 Nano, not a new base model.

Outputs 3

Cosmos 3

model
Architecture MOE

Variants

Name Parameters Notes
Cosmos 3 Super — —
Cosmos 3 Nano — —

Cosmos 3 Technical Report

paper

Cosmos3-DROID

dataset
Size 707GB (71,907 episodes / 22M frames)
License OpenMDW-1.1
world-modelroboticsmultimodalopen-weight

Related