Cosmos 3
Open omnimodal world model — reasons, generates, and acts in one pass.
About
Cosmos 3, unveiled by NVIDIA at GTC Taipei on June 1, 2026, is the first fully open omnimodal world foundation model for physical AI. A single Mixture-of-Transformers architecture unifies what earlier Cosmos releases split across families: world generation (Predict), controlled generation (Transfer), scene understanding (Reason), and action generation (Policy). It ships in Nano (16B, workstation-class) and Super (64B, datacenter) sizes plus a 4B Edge variant for on-device Jetson Thor deployment, with open BF16 weights on Hugging Face under the Linux Foundation's OpenMDW-1.1 license permitting commercial use.
Limitations: the license requires 'Built on NVIDIA Cosmos' attribution and is not OSI-approved open source; the preview-gated build.nvidia.com access was rocky at launch. Benchmarks like the #1 VANTAGE-Bench claim for the Edge variant are NVIDIA-reported. Deployment is smoothest on NVIDIA hardware via NIM microservices.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 9/10
Scored 9/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 63; Trending score 90/100 from researcher assessment.