Google DeepMind Open-Sources WeatherNext Cyclone Forecasting Models

Google DeepMind published new results for WeatherNext Cyclones on August 6, 2026, alongside a Nature paper and an open-source release of model code and weights.

The system is designed to forecast a tropical cyclone's track, intensity and wind structure in a single AI model. DeepMind reports that, across historical 2023-2024 cyclones used for evaluation, WeatherNext Cyclones gained more than 24 hours of lead-time advantage on average across those metrics compared with leading weather models. The company describes its three-day forecasts as comparable in accuracy to what previous models provided at two days.

WeatherNext Cyclones can produce forecasts up to 15 days ahead. DeepMind says a single forecast can be generated in under a minute on a TPU, while a 1,000-member ensemble is used to represent uncertainty and surface lower-probability but consequential scenarios such as rapid intensification.

The release is especially useful to researchers because DeepMind has made WeatherNext 2 and WeatherNext Cyclones code and model weights openly available. It also released WeatherNext 2-mini, a compact variant that can run on a single TPU through a free public Colab notebook.

DeepMind says the research was developed with collaborators including Google Research, the U.S. National Hurricane Center, the Cooperative Institute for Research in the Atmosphere and the UK Met Office. The model was also used operationally during the 2025 hurricane season.

The practical significance is not that AI replaces meteorological agencies. Faster ensemble generation and longer useful lead times can give professional forecasters another high-value signal for preparedness and risk analysis. DeepMind explicitly advises people to rely on their local meteorological agency or national weather service for official forecasts and warnings.

Why this matters for AI and science

Weather forecasting is becoming one of the clearest examples of AI moving from benchmark performance into scientific and operational workflows. The open-source release lowers the barrier for universities, weather agencies and nonprofit researchers to test the model, localize it and compare it with physics-based and other AI forecasting systems.

For OpportunityAtlas readers interested in AI research, climate technology, scientific computing and open models, the release is also a useful signal to watch for new research roles, grants and fellowships at the intersection of machine learning, Earth science and disaster resilience.