Cosmos 3
NVIDIA’s open omni-modal world foundation model for physical AI.
Cosmos 3 is NVIDIA’s open world foundation model for robotics, autonomous vehicles and vision agents. A mixture-of-transformers design pairs a reasoning transformer with a generation transformer so one model can understand and generate text, images, video, ambient sound and actions — for simulation, synthetic data and as a backbone for robot policies. Ships as Super (highest fidelity) and Nano (fast) tiers, with Edge to follow.
- Provider
- NVIDIA
- Type
- World model
- Released
- Jun 1, 2026
- Price
- Open weights; hosted via NVIDIA NIM and partners
- Input
- text, image, video, audio
- Output
- text, video, action
- Open weights
- Yes
Best for
- Multimodal
- Agents
- Research
Strengths
- Open weights on Hugging Face
- Understands and generates actions
- NIM deployment
More from NVIDIA
- Nemotron 3 Super — NVIDIA. Open hybrid Mamba-Transformer MoE for multi-agent applications.
- Nemotron 3.5 Lightning — NVIDIA. Tiny open 30B-A3B model with the lowest first-token latency in the index.
- Nemotron 3 Ultra — NVIDIA. NVIDIA’s largest open model: hybrid Mamba-Transformer MoE, 55B active.
- Nemotron 3 Embed 1B — NVIDIA. Small open embedding model for high-throughput enterprise retrieval.
- Isaac GR00T N1.7 — NVIDIA. Open 3B cross-embodiment VLA for humanoids and manipulators.
- Nemotron 3.5 Content Safety — NVIDIA. Compact 4B multimodal guardrail model for LLM and VLM traffic.
Tools that use it
- NVIDIA NIM — NVIDIA. Prebuilt, optimised inference microservices for NVIDIA GPUs.
- Isaac Lab — NVIDIA. GPU-accelerated robot learning framework on Isaac Sim.