Nvidia Cosmos 3 Edge (huggingface.co)

Nvidia has announced the release of Cosmos 3 Edge, a groundbreaking 4-billion-parameter open world model designed for edge devices like factories, warehouses, and hospitals. This model enables robots and AI agents to comprehend their environments, make real-time decisions, and execute actions efficiently. With optimal performance on NVIDIA GPUs, including the newly launched Jetson T2000 and T3000 modules, Cosmos 3 Edge excels in memory-efficient, high-throughput inference. It stands out as the top performer in vision analytics and robot policy learning within its parameter class, setting new standards for robotics and smart infrastructure. The model features a unique dual-tower architecture: an autoregressive tower for reasoning across language and vision tokens, and a diffusion tower for action predictions and simulations. This design allows the model to maintain a shared representation of the environment, enabling it to predict outcomes and inform decision-making processes based on the dynamics of the world. Developers can fine-tune Cosmos 3 Edge for specific applications using reference post-trained checkpoints and recipes, and benefit from the recently introduced Cosmos 3 Super 4-Step Distillation, which significantly speeds up inference while maintaining quality. Ultimately, Cosmos 3 Edge empowers developers to create customized AI solutions that effectively bridge understanding, simulation, and action in real-world scenarios.