NVIDIA Cosmos 3 將物理 AI 統一為一個 32B 全能模型
nvidia 發布了 cosmos 3，這是一個 320 億參數的混合 transformer 模型，將世界生成、推理和策略任務統一到一個單一架構中。這使得開發人員能夠構建用於機器人和自動駕駛的物理 ai 系統，理解物理學和因果關係，而無需管理多個獨立模型。
nvidia introduced cosmos 3, the first open omni-model foundation for physical ai, shifting from separate models for prediction, transfer, and policy to a single architecture built on a mixture-of-transformers backbone. by processing text, image, video, audio, and action inputs within shared parameter sets that handle both autoregressive reasoning and diffusion generation, the model can simulate physical worlds, predict outcomes, and generate robot policies simultaneously. available in 8b and 32b parameter sizes on hugging face, the model is optimized for generating synthetic data for robotics and autonomous vehicles. this release significantly reduces the complexity of training physical ai systems, enabling developers to deploy physically accurate simulations on hardware ranging from rtx pro 6000 gpus to nvidia hopper and blackwell clusters.