Google DeepMind has introduced an AI model called WeatherNext that predicts cyclones one day earlier with the same accuracy as before. According to a study published in the journal Nature, three-day forecasts now achieve the precision that earlier models only provided after two days. DeepMind makes the code and model weights publicly available.
Model reduces errors in track and storm intensity
The system is based on so-called Functional Generative Networks and generates an ensemble of 1,000 possible weather scenarios for each forecast to also represent rare, risky scenarios. WeatherNext was trained with about 20 terabytes of global weather data and the historical storm database IBTrACS, which includes about 5,000 past cyclones. Despite a comparatively coarse resolution of 28 by 28 kilometers, about a hundred times coarser than classical intensity models, the system reportedly outperforms the error values of established physical models like ECMWF-ENS and HWRF in terms of track, intensity, and wind structure, according to DeepMind. The model calculates a complete fifteen-day forecast in under a minute on a Tensor Processing Unit, significantly faster than classical simulations run on supercomputers. DeepMind estimates the leap as “about a decade” of meteorological progress measured against the developments of the past twenty years, a manufacturer’s assessment that has not yet been independently confirmed. External experts have not yet independently verified the figures. The data basis for the comparison consisted of operational forecasts from the years 2023 to 2025, against which WeatherNext reportedly performed consistently better than the established reference models.
Hurricane Center tests system for the first time with Melissa
The U.S. National Hurricane Center used an early version of WeatherNext operationally during the 2025 Atlantic hurricane season. For Hurricane Melissa, which rapidly gained strength and struck Jamaica, the model provided early indications of the storm’s rapid intensification, according to the center. This additional lead time is said to have given on-site responders more time for evacuations and protective measures; however, DeepMind does not provide specific figures on saved damages or evacuated individuals. In addition to the Hurricane Center, the development involved the Cooperative Institute for Research in the Atmosphere and the UK Met Office. The ongoing 2026 Atlantic hurricane season now offers a first opportunity to test the forecasts of the now more mature model again under real conditions and compare them with the previous laboratory values. A broader version called WeatherNext 2, which covers not only cyclones but also general weather patterns, has reportedly been in operational testing with partner agencies since October 2025.
Google releases model and code openly accessible
DeepMind releases code and trained weights in three variants on GitHub: the complete WeatherNext Cyclones, the broader WeatherNext 2, and the slim version WeatherNext 2-mini, which can reportedly be run in a free Google Colab notebook according to the repository. The code is licensed under the Apache 2.0 license, while the other materials such as the model weights are under CC BY 4.0; both licenses allow free use, redistribution, and adaptation for commercial purposes. Those who do not want to operate their own computing power can also access forecasts directly through Google Earth AI’s interactive web interface Weather Lab, free of charge and without installation. However, productive use of the full model variants requires access to a Tensor Processing Unit or a powerful graphics card; DeepMind does not specify minimum requirements for standard office computers. In addition to the repository and the Colab notebook, DeepMind publishes accompanying documentation on training data and model architecture so that external research groups can independently verify the results and apply them to regions other than the Atlantic, such as the Pacific or the Indian Ocean, where cyclones occur under different names like typhoon or cyclone.
It remains to be seen whether national weather services outside the U.S. will also operationally use the model in the coming months or will initially stick to their own physical simulations. It will also be crucial whether the error values measured in the laboratory can be confirmed over several hurricane seasons; a single practical test with Melissa is not sufficient for that.
