DeepMind Open-Sources WeatherNext, Claiming Forecasts a Full Day Earlier
WeatherNext runs 1,000 ensemble members covering rare high-risk events like rapid intensification — and produces a 15-day forecast in under a minute on one TPU.
DeepMind has open-sourced its weather forecasting model WeatherNext, alongside a claim backed by a paper in Nature: its forecasts are accurate a full day earlier than the previous state of the art.
What the model does
WeatherNext runs a 1,000-member ensemble — meaning it predicts a thousand possible futures for each cyclone instead of one. That breadth matters most for exactly the events that matter most: low-probability, high-impact situations like rapid storm intensification, where a single deterministic forecast can miss the danger entirely.
The accuracy claim is striking: three-day forecasts are as good as what prior models could only provide two days out — roughly a decade's worth of meteorological progress in one step. The model also produces forecasts up to 15 days in advance.
It already saved time in a real hurricane season
This is not a paper-only result. During the 2025 hurricane season, the model helped the US National Hurricane Center make a historic forecast for Hurricane Melissa — predicting its rapid intensification and landfall in Jamaica early enough for advance warnings. The work brought together DeepMind and Google Research engineers with forecasters at the NHC, CIRA and the UK Met Office.
Why open-sourcing it changes things
Weather models are a public good with a direct line to saving lives: better early warnings buy people time to move. DeepMind is now open-sourcing the WeatherNext 2 and WeatherNext Cyclones models used during hurricane season, with code on GitHub — letting national meteorological agencies, researchers and startups build on it without a partnership with Google.
The caveats
Open weights are not the same as operational readiness. Real forecasting agencies will still want to validate the model against their own regions and standards before trusting it with public warnings — and they should. The point is that the validation work, the fine-tuning and the life-saving applications are now things anyone can do.
The best use of AI is not another chatbot. It is a thousand futures of the weather, computed before the storm.