Cosmos 3
Released May 31, 2026
NVIDIA's open omnimodal world foundation model family for physical AI that reasons about and generates text, images, video, ambient sound and actions, usable as a world/video model, VLM, or backbone for robot policies.
- Output
- Simulation video
- Interactive
- No
- Inputs
- text, image, video, actions
- Max session
- —
- Resolution
- 720p
- Frame rate
- —
- Access
- Generally available
- Licence
- OpenMDW 1.1
Strengths & weaknesses
Strengths
- Open weights under a commercial-use licence (OpenMDW 1.1)
- Action-conditioned forward-dynamics generation as well as inverse-dynamics and policy modes
- Three sizes: Edge (4B) for on-device real-time robotics, Nano (16B) for workstation GPUs, Super (64B) for large-scale synthetic data
- NVIDIA reports first place among open models on Physics-IQ and RoboArena
Weaknesses
- Not a real-time interactive generator; produces clips of 5-400 frames
Announced at GTC Taipei; press release and model cards date it 2026-05-31 (some coverage says June 1). Output resolution/fps follow the input request; 720p is the maximum listed input resolution and the examples use 1280x720 at 24 fps, so fps left null. Clip length 5-400 frames (default 189); maxSessionSec left null as no duration is stated (400 frames ≈ 16.7 s at 24 fps, my arithmetic). Super is listed as 64B in NVIDIA's HF blog and 65B on the HF page. Audio input supported (max 0.5 s) — not in the inputs enum. Cosmos 3 Edge (4B, Nemotron-based 2B reasoner per NVIDIA's developer blog) released 2026-07-20 per its HF model card (press dates the announcement 07-16); the Cosmos 3 technical report (06-22) describes Edge as a dense 2B model 'in a later release', conflicting with the card's 4B. A separate Cosmos3-Edge-Policy-DROID robot policy is not part of this entry. OpenMDW SPDX id not confirmed, so spdx null.
Links
API model strings
- huggingface
nvidia/Cosmos3-Nano - huggingface
nvidia/Cosmos3-Super - huggingface
nvidia/Cosmos3-Edge