WeatherNext: DeepMind opens its cyclone forecasting model and gains 24 hours of warning

Published in Nature at the beginning of August, the model provides an accuracy of two days that was previously achieved at three days, and its code and weights are released under the Apache 2.0 license.

On August 6, 2026, Google DeepMind published work in Nature describing the prediction of the trajectory, intensity and structure of tropical cyclone winds by its WeatherNext model. The practical contribution is direct: an additional warning day, a three-day forecast reaching the accuracy previously obtained at two days.

The model was trained on 20 terabytes of meteorological data and on the IBTrACS database of past storms. Unlike traditional physical systems that require fine-mesh inputs, WeatherNext Cyclones operates on a coarser input resolution, on the order of 28 kilometers, which considerably lightens the calculation. The comparisons were conducted on the 2023 and 2024 storms against ECMWF-ENS, the operational reference for the trajectory, and HWRF for intensity.

The code and weights are released on GitHub under the Apache 2.0 license, in three variants: WeatherNext Cyclones, WeatherNext 2 and WeatherNext 2-mini. This openness is far from anecdotal for the national meteorological services of the most exposed countries, which do not always have the means to run the physical reference models.