Google DeepMind’s WeatherNext storm model can predict cyclones with about a day more lead time than leading operational systems, according to a paper published in Nature on August 6. Its three-day forecast matches what previous models delivered at two days, a gain researchers compare to a decade of meteorological progress.
WeatherNext Cyclones produces ensembles of track, intensity, and size predictions up to 15 days out, scaling to 1,000 scenarios per storm, twenty times the earlier system’s runs. During the 2025 Atlantic season it ran alongside National Hurricane Center operations and helped forecasters anticipate Hurricane Melissa’s rapid intensification and Jamaica landfall, marking the first time the center predicted a Category 5 storm while it was still Category 1.
The surprise finding: the model runs on 28-by-28-kilometer inputs, roughly 100 times coarser than traditional regional models, and a mini variant at 111 kilometers still performs well. The researchers say they do not yet understand how it extracts intensity signal from coarse data, calling it an open question.
Google is open-sourcing WeatherNext Cyclones, WeatherNext 2, and WeatherNext 2-mini on GitHub, with the compact version running on a single TPU in a free Colab notebook. A 15-day forecast takes under a minute on one TPU. Forecasters stress the human element still matters, noting a hurricane is more than a track and intensity readout.
