On June 1, 2026, at the GTC conference held in Taipei,NVIDIA officially releases Cosmos 3——The world’s first fully open “Omnimodal World Model”.
core competencies
The core breakthrough of Cosmos 3 lies in itsMixture-of-Transformers——Pairing inference Transformers with expert generation Transformers allows the model to understand object interactions, motion, and spatiotemporal relationships before generation.
Cosmos 3 can play four roles simultaneously:
- visual language model: Understand physical scenes and answer questions about them
- video generator: Generate physically realistic videos from text or images
- world simulator: Predict how objects will move and interact - can be used to train robots
- strategy model:Planning action paths for robots and self-driving cars
In terms of benchmark testing, Cosmos 3 ranks first among all open source models, covering multiple physical AI benchmarks such as Artificial Analysis, Physics-IQ, PAI-Bench, and R-Bench.
Cosmos Alliance and Open Ecosystem
Also released with Cosmos 3NVIDIA Cosmos Coalition——One by Agile Robots, Black Forest Labs,Runway, Skild AI, etc. to jointly promote the development of the open world model.
NVIDIA also updated the Cosmos platform, adding new data sets in the fields of robotics, physics, human motion, autonomous driving, warehousing security and spatial reasoning, as well as physical AI agent skills such as neural scene reconstruction, defect image generation and video enhancement.
User experience and limitations
For robot developers and autonomous driving engineers, the greatest value of Cosmos 3 isCompress the training and evaluation cycle of physics AI from months to days.
But we must also clearly realize: Cosmos 3 is a developer tool.Not a product for ordinary consumers.R&D accelerationrather than direct deployment.
Overall Score
| 维度 | Score | evaluate |
|---|---|---|
| functional completeness | 9.0 / 10 | Full modal support (text + image + video + audio + action), integrated reasoning + generation, extremely wide coverage |
| 易用性 | 6.5 / 10 | Positioned as a low-level tool for developers, the deployment and debugging threshold is high and requires the support of a professional engineering team. |
| Cost-effectiveness | 8.0 / 10 | Open source weights + DGX Cloud save a lot of costs compared to training from scratch; but the hardware investment is still not small |
| 中文支持 | 7.0 / 10 | Graphic and text understanding supports Chinese, but the physics AI community and documentation are mainly in English, and Chinese materials are scarce. |
| 输出质量 | 8.5 / 10 | The physical accuracy and benchmark scores are both No. 1 in open source, but there are still gaps in real scenarios. |
Overall rating: 7.8/10(Extremely valuable to robot/autonomous driving developers; not applicable to ordinary users)
[AI Dash](https://aidashxp.com) — Discover the best AI tools