Accurately predicting Earth system dynamics across various spatiotemporal scales is crucial for sustaining socio-ecological systems. While traditional numerical models provide detailed physics, they are computationally intensive and struggle to integrate widely available multiscale data. Problem-specific machine learning models, although excellent for single tasks, lack the versatility needed for broader applications. To overcome these limitations, we are developing ORBIT, an AI foundation model designed to enhance Earth system modeling. ORBIT comes in four sizes, with capacities of 115 million, 1 billion, 10 billion, and 113 billion parameters. Each model incorporates 91 atmospheric variables and is pre-trained on ten CMIP6 datasets. Once pre-trained, ORBIT can be fine-tuned for a broad array of Earth system modeling tasks, including weather forecasting and climate projection. Our evaluations indicate that even the smallest ORBIT model, with 115 million parameters, is competitive with both the state-of-the-art AI foundation model, specialized machine learning models, and the operational Integrated Forecasting System for weather predictions up to 30 days ahead. Moreover, ORBIT’s performance notably improves with increasing model size, with the 113 billion parameter model demonstrating superior outcomes. Future work will focus on further development of ORBIT and expanding its capabilities for additional Earth system modeling tasks.