NVIDIA Unveils 4B-Scale World Model 'Cosmos 3 Edge' for 'On-Device Physical AI'

NVIDIA introduced a new world model for edge computing environments, enabling real-time understanding and inference of surrounding environments to generate robot actions, thereby strengthening its physical AI domain. The company unveiled the world model 'Cosmos 3 Edge' for robots and vision AI agents on 15th (local time). As part of the open-source Cosmos 3 series, this new model is built on NVIDIA's flagship model, 'Nemotron,' with 40 billion parameters. It supports robots and vision AI agents to recognize their surroundings, make real-time decisions, and generate robot actions directly on NVIDIA edge computers. Unlike previous large models centered on the cloud, Cosmos 3 Edge is designed to perform inference and control within the device itself, minimizing network latency and enabling immediate decision-making in physical AI environments such as manufacturing, logistics, autonomous driving, and industrial automation. Developers can quickly customize the model to specific robots, vehicles, sensors, and operational environments using the open-source NVIDIA Cosmos framework within a day. Its lightweight structure allows execution on edge GPUs, and it's designed to rapidly perform post-training for task-optimized world action models. Cosmos 3 Edge runs on NVIDIA RTX GPUs, DGX systems, and Jetson platforms, including newly released 'Jetson T2000' and 'T3000' modules. This model extends NVIDIA's Cosmos 3 world model platform launched in June, enhancing learning data generation and action prediction technologies for robots and autonomous systems.