DeepMind's cyclone model buys forecasters an extra day, and it's open source
Published in Nature with the National Hurricane Center and the UK Met Office: three day cyclone forecasts as accurate as the old two day ones, a thousand scenarios per storm, and the model on GitHub.
A three day cyclone forecast now matches what two day forecasts used to be. The model is on GitHub.
Google DeepMind and Google Research published WeatherNext in Nature on August 6, a cyclone model built with the US National Hurricane Center, the UK Met Office, and CIRA. The headline claim: its three day forecasts for a cyclone's track and intensity are as accurate as the two day forecasts from prior systems. Forecasters call that an extra day, and evacuation planners measure it in lives.
What the model actually does
WeatherNext generates 1,000 possible scenarios per cyclone and predicts track, intensity, and wind structure in a single model, where traditional pipelines split those jobs across separate systems. It runs a 15 day forecast in under a minute on a TPU. The training data is about 20 terabytes of atmospheric history plus roughly 5,000 storms from the IBTrACS database.
The counterintuitive part
The model works at a 28 by 28 kilometer resolution, around 100 times coarser than the physics simulations it is compared against. It gains accuracy anyway, the same pattern GraphCast showed for general weather: learned models extract more signal per grid cell than numerical simulation gets from brute resolution. It already ran through the 2025 hurricane season, including Hurricane Melissa's rapid intensification ahead of its Jamaica landfall.
Why a build studio cares
Two reasons. First, the release model: this is a Nature-reviewed system, deployed with the agencies that issue real warnings, and then open-sourced on GitHub rather than parked behind an API. That combination is still rare and worth noticing every time it happens. Second, the shape of the win: a domain where the incumbent is an expensive simulation, beaten by a learned model that is 100x coarser and 1,000x faster, is a shape we look for in client problems too. The tradeoff is scope: WeatherNext is a cyclone specialist, and a decade-of-progress claim for one storm class does not make general forecasting solved.
Next step: read DeepMind's announcement and The Decoder's summary, or poke at the scenarios yourself in Weather Lab. If you have a simulation-shaped problem that might be a model-shaped problem, write to us at hello@gattyworks.com.