DeepMind Open-Sources WeatherNext Cyclones Under Apache 2.0 After a Nature Paper Showing a Full Extra Day of Warning

Google DeepMind·high signalXBlueskyLinkedInCopy link

Google DeepMind published in Nature on August 6 that WeatherNext Cyclones matches at three days the track, intensity and wind-structure accuracy previously achievable only at two days — more than 24 hours of added lead time. The model was trained on 20 TB of weather data plus the IBTrACS historical storm database and uses Functional Generative Networks to sample 1,000 possible forecasts, which is what lets it catch rapid intensification. Code and weights are on GitHub under Apache 2.0 in three variants — WeatherNext Cyclones, WeatherNext 2, and a WeatherNext 2-mini that runs in a free Colab notebook — and the National Hurricane Center already used it operationally on Hurricane Melissa's Jamaica landfall.