NVIDIA has announced the global availability of Earth-2, a comprehensive family of fully open-source AI models designed for weather and climate forecasting. The platform provides a complete software stack covering every stage of the weather prediction process, from data processing to high-resolution visualization. By shifting from resource-intensive physics models to an end-to-end AI pipeline, Earth-2 enables researchers, startups, and government agencies to run, fine-tune, and deploy sophisticated forecasting systems on their own infrastructure with unprecedented speed and cost-efficiency. The models are accessible via Hugging Face and can be used with Earth2Studio, an open-source Python ecosystem.
The Earth-2 model family consists of three primary components. Earth-2 Medium Range, powered by the Atlas architecture, delivers high-accuracy predictions for over 70 weather variables up to 15 days in advance, outperforming leading open models like GenCast on key industry benchmarks. Earth-2 Nowcasting, using the StormScope architecture and generative AI, predicts satellite and radar imagery to transform country-scale forecasts into kilometer-resolution, zero- to six-hour predictions of local storms and hazardous weather in just minutes. This model represents the first AI system to outperform traditional physics-based weather prediction models in short-term precipitation forecasting by simulating storm dynamics directly.
Coming soon to the platform is Earth-2 Global Data Assimilation, powered by the HealDA architecture, which produces initial conditions for weather prediction. The model generates snapshots of the current atmosphere—including temperature, wind speed, humidity, and air pressure at thousands of locations around the globe—in seconds on GPUs instead of hours on supercomputers. When coupled with Earth-2 Medium Range, this combination delivers the most skillful forecasting predictions produced by an open, entirely AI pipeline. Additionally, the CorrDiff model enables generative AI downscaling 500 times faster with a 10,000-fold improvement in energy efficiency, while FourCastNet 3 accelerates AI-based global weather forecasting with datasets up to 20 times larger.
The platform is already being adopted by leading organizations across multiple industries, from energy and finance to global weather agencies. NVIDIA Earth-2 is powering a new era of weather intelligence, with organizations leveraging these open models and tools to transform raw data into actionable insights. The platform offers interactive workflows ranging from city-scale to global-scale simulations, integrating with Google Photoreal Tiles to address urban environment questions. NVIDIA is collaborating with companies like MITRE and Tomorrow.io to advance next-generation extreme weather simulation, and the platform's Fourier Neural Operator combined with PhysicsNeMo accelerates carbon capture and storage analysis by 700,000 times. The company demonstrated that FourCastNet can generate 21-day weather trajectories of 1,000 ensemble members in one-tenth the time previously required for a single ensemble, with 1,000 times less energy consumption.