NVIDIA opens up Earth-2, a free weather-AI toolkit, as public forecasting funding tightens

NVIDIA released Earth-2, which it describes as the first fully open and accelerated software stack for weather and climate AI, at the American Meteorological Society's annual meeting on January 26, 2026. The package bundles pretrained models, software libraries, customization recipes and inference tools so that weather services, researchers and companies can build forecasting systems without assembling the pieces themselves. The models are available through public repositories on GitHub and Hugging Face.

The release includes several distinct tools. An Earth-2 medium-range model forecasts more than 70 weather variables up to 15 days out; a nowcasting model uses generative AI trained on satellite and radar data to predict cloud and rainfall systems at kilometer scale for the next zero to six hours; and a data-assimilation model can generate the atmospheric starting conditions a forecast needs in seconds, a step that traditionally consumes hours on supercomputers. NVIDIA says it is the first time AI has surpassed traditional physics-based models for short-term precipitation in this setup.

“The package bundles pretrained models, software libraries, customization recipes and inference tools so that weather services, researchers and companies can build forecasting systems without assembling the pieces themselves.”

The open approach is the most consequential part. By publishing the stack freely, NVIDIA lowers the cost for national weather agencies and university groups that cannot afford to develop such systems alone, and it folds in models from partners including the European Centre for Medium-Range Weather Forecasts, Microsoft and Google. Early adopters range from the U.S. National Weather Service to the Israel Meteorological Service and Taiwan's Central Weather Administration.

There are caveats. AI weather models still depend on the quality of observational data and on traditional physics for many tasks, and "open" tools require technical skill and computing power to run well. A forecast also only helps if it reaches people who can act on it. But as some public forecasting budgets come under pressure, a free, shared toolkit that puts advanced weather prediction within reach of more countries is a practical step toward better warnings for the communities that need them most.