Google DeepMind said on August 6, 2026 that Nature had published its work on WeatherNext, which the company reports reaching state-of-the-art accuracy on a cyclone’s track, intensity, and wind structure — and released the code and model weights at the same time 1. The gain amounts to roughly a day, by DeepMind’s account: forecasts three days out now reach the quality that earlier systems managed only two days out, on average 1. Benchmarked against other leading weather models on past cyclones, the company reports an average advantage in lead time across track, intensity, and wind structure of more than 24 hours 1. DeepMind puts that scale of change at about a decade of progress in meteorology 1.
An operational example is cited from the 2025 hurricane season, when DeepMind says the model helped the U.S. National Hurricane Center (NHC) anticipate Hurricane Melissa’s rapid intensification and its landfall in Jamaica, making an early warning possible 1. The NHC, the Cooperative Institute for Research in the Atmosphere (CIRA), the UK Met Office, and weather agencies elsewhere took part in the work 1.
One technical point stands out. WeatherNext Cyclones works from data at 28x28km — a hundred times coarser than conventional models — and DeepMind says why accuracy survives at that scale is still unresolved 1. A 15-day run finishes in under a minute on a TPU, and the ensemble has grown from 50 members last year to 1,000 this year 1. Three models are going out: WeatherNext Cyclones, which ran through the hurricane season; the later WeatherNext 2; and WeatherNext 2-mini, small enough for one TPU inside a free public Colab notebook 1. DeepMind adds that official forecasts and warnings should still come from local meteorological agencies 1.
Sources
- WeatherNext: AI model achieves breakthrough in forecasting cyclones - Google DeepMind official blog (August 6, 2026)