Cosmos 3 Super

Spec

Attribute	Value
Developer	NVIDIA
Released	2026-06-01
Announced	2026-06-01
Context window	unknown
Pricing	unknown
License	MIT (open weights)
Availability	HuggingFace — open weights (nvidia/Cosmos3-Super)
Parameters	64B (32B reasoner + 32B generator)
Architecture	Mixture-of-Transformers (MoT)
Supported HW	NVIDIA Hopper, Blackwell
HuggingFace	nvidia/Cosmos3-Super

Release Date

2026-06-01 (Cosmos 3 series launched together: Super / Nano / Edge pending)

What Is It

The world's first fully open omnimodal physical AI foundation model. It generates and reasons over text, image, video, audio, and action within a single unified model.

Mixture-of-Transformers (MoT) architecture:

• Autoregressive tower: discrete token generation (text, action commands)
• Diffusion tower: continuous multimodal generation (image, video, audio)

The two towers cooperate within a single inference cycle — unifying world understanding (reasoning), world generation (world gen), and action prediction (action).

Benchmarks

• R-Bench: #1 among open models (physical AI reasoning)
• Artificial Analysis: #1 open model (physical AI category)

Training Data

• 20 trillion tokens of multimodal data
• ~1 billion images
• 400 million real + synthetic videos
• Ambient audio, text, human/robot action data

Use Cases

• Synthetic data generation for robot training (Synthetic Data Generation)
• Autonomous driving simulation
• Post-training foundation for physical AI models
• Virtual validation of industrial automation scenarios

Model Family

Model	Size	Purpose
Cosmos 3 Super	64B	SDG, highest physical accuracy
Cosmos 3 Nano	Small	Real-time inference, fractions of a second
Cosmos 3 Edge	TBD (coming soon)	Real-time inference on edge devices

Cosmos Coalition

A global collaborative consortium launched by NVIDIA at the same time:

• Agile Robots, Black Forest Labs, Generalist, LTX, Runway, Skild AI

Significance

1. Omnimodel unification: world generation + reasoning + action in a single model — previously this required separate pipelines
2. MIT-licensed open release: free for research/commercial use → rapid community development expected
3. 20T multimodal training: LLM-scale data volume applied to physical AI for the first time
4. Cosmos Coalition: an ecosystem lock-in strategy — an attempt to secure a foothold across the robotics/AV industry

Open Questions

• Real-world robot sim-to-real performance validation (community reproduction needed shortly after release)
• Edge model release timing
• Against the competition: Gemini Robotics ER 1.6 vs Cosmos 3 real-world performance comparison

Sources

• (source)
• HuggingFace blog: https://huggingface.co/blog/nvidia/cosmos-3-for-physical-ai
• GlobeNewswire (official): https://www.globenewswire.com/news-release/2026/06/01/3303987/0/en/NVIDIA-Launches-Cosmos-3-the-Open-Frontier-Foundation-Model-for-Physical-AI.html

Related

• NVIDIA
• Jim Fan
• Embodied Agents
• Gemini Robotics ER 1.6 — DeepMind competing physical AI model