NVIDIA has unveiled three new open-source models as part of its NVIDIA Earth-2 initiative, significantly advancing the field of AI-powered weather and climate forecasting. These models are designed to streamline and enhance capabilities across the entire weather prediction stack, including data assimilation, forecasting, nowcasting, and downscaling. Built on NVIDIA’s accelerated AI infrastructure, Earth-2 offers developers a fully open, customizable platform to create sovereign, high-precision weather and climate simulations tailored to their specific needs using their own data and hardware. The first model, Earth-2 Nowcasting, introduces StormScope—a novel generative AI architecture that delivers kilometer-scale, zero- to six-hour forecasts of severe local storms and hazardous weather. Trained on geostationary satellite data from GOES over the contiguous United States, it generates high-resolution predictions in minutes, outperforming traditional physics-based models in short-term precipitation forecasting. By directly modeling storm dynamics and predicting satellite and radar observations, StormScope enables faster, more accurate local forecasts critical for disaster preparedness. The second model, Earth-2 Medium Range, leverages a new architecture called Atlas to produce highly accurate 15-day global forecasts across more than 70 atmospheric variables such as temperature, wind, humidity, and pressure. Using a latent diffusion transformer, it predicts incremental atmospheric changes while preserving key physical structures, minimizing errors over time. On industry-standard benchmarks, Earth-2 Medium Range surpasses leading open models like GenCast, demonstrating its superior skill in probabilistic medium-range forecasting. The third model, Earth-2 Global Data Assimilation, is set to launch soon on Hugging Face. Powered by HealDA, it rapidly generates initial atmospheric conditions—such as temperature, wind speed, and pressure—across thousands of global locations in seconds using GPUs, compared to hours on traditional supercomputers. When integrated with Earth-2 Medium Range, this end-to-end AI pipeline produces the most accurate open-source forecasts to date, significantly reducing the time and computational cost of initializing weather models. These new models join a growing suite of open-source Earth-2 tools, including FourcastNet3, CorrDiff, cBottle, and DLESym, forming a comprehensive ecosystem for AI-driven climate science. Developers can now use NVIDIA Earth2Studio, an open-source Python framework, to quickly build and deploy inference pipelines for weather and climate simulations. Earth2Studio provides ready-to-use tools and workflows, enabling rapid prototyping and deployment of AI models directly on Hugging Face. The open nature of Earth-2 allows organizations to maintain full ownership and control over their predictions, supporting data sovereignty and regulatory compliance. This is especially valuable for governments, research institutions, and private sector entities seeking to build resilient, localized forecasting systems without relying on proprietary platforms. NVIDIA’s approach combines cutting-edge AI with domain-specific physics, aiming to bridge the gap between data-driven modeling and traditional meteorology. The research behind these models, published in peer-reviewed papers, underscores the importance of accurate initial conditions and the transformative potential of AI in improving forecast skill. With these advancements, NVIDIA is empowering a new generation of climate and weather scientists, engineers, and developers to build faster, more accurate, and customizable forecasting systems. By democratizing access to high-performance AI tools, Earth-2 is poised to accelerate innovation in climate resilience, disaster response, agriculture, energy planning, and beyond.