DeepMind open-sources WeatherNext cyclone forecasting model
A Nature paper puts Google DeepMind's WeatherNext Cyclones ahead of operational systems by over a day of lead time, with code and weights released on GitHub.
Google DeepMind published its WeatherNext cyclone forecasting work in Nature on Thursday and simultaneously released the model's code and weights on GitHub, making a system already used inside operational hurricane forecasting freely available to anyone.
The result
WeatherNext Cyclones produces probabilistic 15-day forecasts of a storm's track, intensity and wind structure. DeepMind reports state-of-the-art accuracy across all three, and frames the gain in the currency forecasters actually use: lead time. On average, its three-day predictions match the quality that prior systems delivered at two days — better than a full extra day of warning. In a field where operational skill has historically improved by roughly a day per decade, that is a decade of progress arriving at once.
The model was trained on years of global atmospheric reanalysis data — on the order of 20 terabytes — combined with the IBTrACS archive of nearly 5,000 historical cyclones, which supplies the storm-specific structure that general atmospheric training alone tends to smooth away. DeepMind cites Hurricane Melissa as a field test: forecasters received early signals of a Category 5 landfall five days ahead with 80 percent confidence.
Release and partners
The code and weights are on GitHub for academic, operational and commercial use, and a compact WeatherNext 2-mini variant runs in a free Colab notebook. Predictions are also served through DeepMind's Weather Lab. The work was done with the US National Hurricane Center, the Cooperative Institute for Research in the Atmosphere, the UK Met Office and other national agencies, and DeepMind says it is expanding access for forecasters this season.
Why it matters
Machine-learning weather models have been outperforming numerical simulation on general forecasting for two years, but tropical cyclones — rare, small-scale, high-consequence — were the hard case that kept operational meteorologists cautious. Publishing in Nature settles the methodological argument in public; releasing the weights settles the access question. Any national meteorological service, including those without the supercomputing budgets that traditional cyclone modelling demands, can now run a frontier system on modest hardware. It is also a rare instance of a frontier lab open-sourcing a model whose primary value is not commercial, and the clearest current example of AI capability translating into measurable public-safety outcomes.