DeepMind's WeatherNext buys forecasters an extra day on cyclones

A weather model buying forecasters back a full day of lead time on a hurricane track is not a small claim. That is roughly what DeepMind says about WeatherNext, arguing its three-day cyclone forecasts are as accurate as the two-day forecasts prior models produced. In the company's framing, that closes about a decade of meteorological progress in one release.

The proof point the post leans on is Hurricane Melissa, which struck Jamaica in October 2025 as the strongest storm on record to make landfall there and tied for the strongest Atlantic hurricane ever recorded. According to DeepMind, WeatherNext flagged a Category 5 Jamaica landfall five days out with 80% confidence, rising to near 100% at three days, running 50 what-if scenarios in the process. The National Hurricane Center's 2025 verification put WeatherNext as the top-performing individual model for both track and intensity, and NHC director Michael Brennan is quoted on how quickly cyclone structure and intensity can shift, which is what makes these storms hard to predict.

The technical shift underneath the accuracy number matters just as much. WeatherNext operates on a 28x28km input grid, roughly 100 times coarser than traditional physics-based models, and produces 1,000-member ensembles in under a minute on a TPU. That reframes cyclone forecasting from a supercomputer workload into something a small national met service could run itself. DeepMind has open-sourced WeatherNext Cyclones, WeatherNext 2 and WeatherNext 2-mini on GitHub, weights included, and has been working with meteorological agencies in the Philippines, Taiwan, Indonesia and Vietnam.

The honest caveat is that these are DeepMind's own numbers on its own headline storm, and the model was trained end-to-end on roughly 20 terabytes of atmospheric data and about 5,000 historical storms from the IBTrACS database, so how it holds up on the kinds of systems climate change is generating outside that training distribution is a genuine open question. What the reporting doesn't get into is how forecasters actually weight the AI output against traditional ensembles when the two disagree on intensity.

For smaller-country met services and for reinsurers pricing cyclone risk, an extra confident day of lead time is the part worth watching.