Google DeepMind open-sources WeatherNext AI model

Google DeepMind and Google Research, in collaboration with the National Hurricane Center (NHC), the UK Met Office, and the Cooperative Institute for Research in the Atmosphere (CIRA), have released the code and weights for their WeatherNext AI models. Published in Nature, the research introduces WeatherNext Cyclones and WeatherNext 2, which improve cyclone track, intensity, and wind structure predictions. On average, the models gain a 24-hour lead time advantage over existing systems, matching the accuracy of a three-day forecast with what older models could only achieve for a two-day outlook. This leap represents roughly ten years of traditional meteorological progress.

The system achieves this by bridging the gap between global atmospheric tracking and localized intensity modeling. Trained on nearly 20 terabytes of global atmospheric data and the IBTrACS database of nearly 5,000 historical storms, the model was evaluated on historical cyclones from 2023 to 2024. WeatherNext uses Functional Generative Networks (FGNs) to generate a single 15-day forecast in under a minute on a Tensor Processing Unit (TPU). While traditional physics models require high spatial resolutions, WeatherNext Cyclones operates at a resolution of 28x28km, which is 100 times coarser than traditional setups. A smaller version, WeatherNext 2-mini, runs at 111x111km resolution and can be executed on a single TPU within a free public Colab notebook.

For practitioners and meteorologists, this release changes how extreme weather risks are calculated. Instead of running small ensembles, the model scales up to generate 1,000 possible scenarios per cyclone, allowing forecasters to map out rare but devastating tail-risks like rapid intensification. The model has already proven its utility in the real world; during the 2025 hurricane season, it helped the NHC predict the rapid intensification and Jamaican landfall of Hurricane Melissa. By making WeatherNext Cyclones, WeatherNext 2, and WeatherNext 2-mini open source, DeepMind enables researchers to run highly accurate, probabilistic weather simulations without needing massive supercomputing infrastructure.