AIBRIEF
A peer reviewed Nature paper and public model weights put the tool in forecasters' hands, with the source itself flagging a track versus intensity trade off the AI has not dissolved.
Google DeepMind has released the model weights for WeatherNext Cyclones, the AI that helped the National Hurricane Center forecast Hurricane Melissa's rapid intensification and Jamaica landfall in October 2025. The release lands alongside a peer-reviewed Nature paper and an arXiv preprint, turning the system from a private benchmark into something forecasters at the NHC, the Cooperative Institute for Research in the Atmosphere (CIRA), the UK Met Office, and equivalent agencies can now inspect, extend, or push back on.
The model predicts 1,000 possible scenarios per storm and, on average, gives forecasters roughly an extra day of predictive lead. Three-day forecasts now match what older systems delivered at two days, Google DeepMind says, for track, intensity, and wind structure. The company frames the jump as "a decade of meteorological progress," a characterization, not an independent benchmark.
Tropical cyclones, hurricanes in the Atlantic and typhoons in the Pacific, have caused more than 700,000 deaths and $1.4 trillion in losses over 50 years, according to Google DeepMind. Timelier warnings change evacuation, sheltering, and grid preparation, but the source itself flags a trade-off: track and intensity predictions pull against each other in traditional models, and the AI version has not dissolved that tension.
The handoff is the news. The weights are public; the compute to run them is not trivial.