Nvidia has launched a new Earth-2 family of open models, libraries and frameworks designed to support AI-based weather and climate forecasting, making advanced weather AI more accessible to organizations worldwide.
The announcement was made at the American Meteorological Society’s annual meeting, where Nvidia described Earth-2 as the world’s first fully open, accelerated software stack covering the full forecasting workflow – from processing observational data to generating medium-range global forecasts and short-term local storm predictions.
Nvidia said the open nature of Earth-2 allows scientists, weather agencies, developers, enterprises and governments to run, fine-tune and deploy production-ready weather AI models on their own infrastructure. The company positioned the platform as a lower-cost, faster alternative to traditional physics-based numerical weather prediction, which typically requires large supercomputing resources.
The Earth-2 stack includes pretrained models, inference libraries and customization tools aimed at reducing computational time while maintaining or improving forecast accuracy. According to Nvidia, AI-powered forecasting can significantly reduce compute requirements, enabling more countries and organizations to develop application-specific forecasting systems.
Three new open models were introduced as part of the Earth-2 family. Earth-2 Medium Range, built on a new architecture called Atlas, supports forecasts up to 15 days ahead across more than 70 weather variables. Earth-2 Nowcasting, powered by the StormScope architecture, generates kilometer-scale, zero- to six-hour forecasts of local storms using generative AI. Earth-2 Global Data Assimilation, based on the HealDA architecture, produces initial atmospheric conditions for forecasts in seconds on GPUs rather than hours on supercomputers.
These models join existing open Earth-2 components, including CorrDiff for high-resolution downscaling and FourCastNet3 for rapid, high-accuracy global forecasting. The platform also integrates open models from organizations such as the European Centre for Medium-Range Weather Forecasts, Microsoft and Google, and supports training via Nvidia’s open-source PhysicsNeMo framework.
Nvidia said Earth-2 models are already being evaluated or used operationally by weather agencies, energy companies and risk and analytics firms. Users include the Israel Meteorological Service, Taiwan’s Central Weather Administration, The Weather Company and the US National Weather Service, as well as energy and grid operators such as TotalEnergies, Eni and Southwest Power Pool.
“The revolution of new AI weather tools for forecasting is very exciting,” said Julian Green, co-founder and CEO of AI weather provider Brightband, which is running Earth-2 Medium Range operationally. “The model being open source speeds up innovation, allowing easier comparison and improvements.”
Earth-2 Medium Range and Nowcasting are now available through Nvidia Earth2Studio, Hugging Face and GitHub, while Earth-2 Global Data Assimilation is expected to be released later in 2026.
In related news, Tomorrow.io announces DeepSky AI-native weather-sensing constellation
