Google DeepMind releases WeatherNext Cyclones code and weights after NHC use

Why it matters

Researchers can now test Google's claimed one-day cyclone forecasting gain against storms and regions outside its published benchmarks, while examining a model whose guidance NHC forecasters used in real operations.

Six Google DeepMind researchers - Ferran Alet, Tom R. Andersson, Ilan Price, Stratis Markou, Andrew El-Kadi and Dominic Masters - are listed as equal-contribution lead authors on the Nature paper describing WeatherNext Cyclones. Google DeepMind released the model's code and weights on August 6, 2026, after forecasters at the US National Hurricane Center used its guidance during the 2025 hurricane season.

The paper has 32 authors from Google DeepMind, Google Research, the NHC, Colorado State University's Cooperative Institute for Research in the Atmosphere, the UK Met Office and the University of Waterloo. Their work tackles the central problems in hurricane forecasting: predicting where a storm will travel, how strong it will become and how far its damaging winds will extend.

In its August 6 announcement, Google DeepMind said WeatherNext Cyclones delivered an average lead-time advantage exceeding 24 hours across storm track, intensity and wind structure. Google DeepMind said its three-day forecasts were about as accurate as the two-day forecasts produced by earlier systems. The reported historical testing described in Google's source article covered cyclones from 2023 and 2024, while a separate graph compared three-day forecast errors from 2023 through 2025.

One model for the whole storm

Traditional cyclone forecasting divides the problem across model types. Large global systems capture the atmospheric currents steering a storm, while computationally expensive regional models resolve the local processes around its core. Those core processes drive intensification, including the sudden jumps in wind speed that leave emergency agencies with little time to prepare.

Alet and his co-authors trained WeatherNext Cyclones across both data domains. The model learned from nearly 20 terabytes of global atmospheric information and the IBTrACS archive of almost 5,000 historical storms.

Google DeepMind says the cyclone system can generate as many as 1,000 scenarios for each storm. A larger ensemble gives forecasters more opportunities to identify low-probability outcomes such as rapid intensification instead of allowing the average prediction to conceal them.

How the NHC has used the guidance

WeatherNext entered the NHC's workflow before the paper and code release. The agency's preliminary report for the 2025 season said 2025 was the first season in which the NHC incorporated AI-based models into real-time operations and described the Google DeepMind guidance, identified as GDMI, as useful. The report also cautioned that some AI systems were still under development and were not consistently delivered in time for routine use.

Hurricane Melissa supplied the highest-stakes example. In October 2025, WeatherNext projected a Category 5 landfall in Jamaica while Melissa was still far weaker. The NHC used the guidance in its analysis of the storm, while Google DeepMind later said the system had predicted Melissa's rapid intensification and Category 5 landfall in Jamaica.

Opening the model creates a harder test

The WeatherNext repository includes WeatherNext 2, WeatherNext Cyclones and related research materials.

Access to the code and weights lowers the barrier for researchers and weather agencies to reproduce the paper's results and compare forecasts across regions. It also exposes Google's performance claims to tests designed and run outside the company.

Google describes WeatherNext as an experimental research project whose outputs cannot replace alerts and warnings from government meteorological agencies. External researchers can now examine how the model handles storms, ocean basins and rare conditions outside the reported averages. Weather models earn confidence over seasons, including the cases where they become outliers or arrive too late to help.

For Alet, Andersson, Price, Markou, El-Kadi, Masters and their collaborators, releasing the weights moves the project into that longer evaluation cycle. NHC documents establish that forecasters used Google DeepMind guidance in real-time operations and during Melissa. They do not independently validate Google's claimed one-day accuracy advantage. Testing by outside researchers will determine how consistently that result holds beyond Google's reported benchmarks.