이미지: NVIDIA
Summary
- In July, NVIDIA signed an open letter on 'open weights' together with more than 200 companies and institutions
- Cosmos 3 is an open physical AI foundation model combining vision reasoning, world generation, and action prediction
- Cosmos 3 is released under the OpenMDW 1.1 license, allowing further training with proprietary data and hardware
- 서한 서명
- 2026년 7월, 'Open Weights and American AI Leadership' 공개서한에 200개 이상 기업·기관 참여
- 모델명
- NVIDIA Cosmos 3, 믹스처 오브 트랜스포머(mixture-of-transformers) 아키텍처 기반
- 라이선스
- 리눅스 파운데이션의 OpenMDW 1.1
- 기능
- 비전 언어 모델, 월드 생성, 행동 예측을 하나의 모델 계열로 통합
- 연계 도구
- NVIDIA Omniverse 라이브러리 및 OpenUSD 프레임워크로 시뮬레이션 환경 구축 지원
NVIDIA said that in July it signed an open letter titled "Open Weights and American AI Leadership" together with more than 200 companies and institutions. The letter argues that AI leadership is determined not by a single frontier model, but by how far an open ecosystem reaches across every industry.
Cosmos 3, an Open Model for Physical AI
As a concrete example of this movement, NVIDIA introduced the open world model 'Cosmos 3.' Built on a mixture-of-transformers architecture, Cosmos 3 is an open physical AI foundation omni-model that combines vision reasoning, world generation, and action prediction within a single model family. Developers can use Cosmos 3 as a vision-language model, or apply it to understand scenes, generate synthetic data, and simulate future states.
Robots, autonomous vehicles, and vision AI systems each operate with different sensors and environments, meaning general-purpose models alone have limits. Bridging that gap requires access to model weights, a license that permits further training, and related tooling, according to NVIDIA. Cosmos 3 is released under the Linux Foundation's OpenMDW 1.1 license, enabling teams to specialize the model using their own data and hardware.
Integration With Omniverse and OpenUSD
NVIDIA noted that model specialization is only part of the workflow, and that building environments for data generation, simulation, and behavior testing is also necessary. To support this, the Omniverse library included in the NVIDIA Agent Toolkit provides the ability to pre-build simulation-ready environments, while OpenUSD serves as an open framework for structuring and reusing complex 3D data across digital twin, simulation, and synthetic data generation workflows.
