Google DeepMind unveils WeatherNext: a day's head start on hurricanes, and an open question about the physics

On 6 August 2026 Google DeepMind and Google Research announced, through a study published in Nature, that their WeatherNext model reaches state-of-the-art results in forecasting the track, intensity and wind structure of tropical cyclones. Publication in Nature is attested by the official announcement and by several outlets: we were not able to find the paper's direct DOI. According to the official release, the average gain is more than 24 hours of lead time, making a three-day forecast as accurate as a two-day forecast was with earlier operational models. In the announcement the company speaks of “roughly a decade of meteorological progress in one model” and cites an improvement of about 100 km in position error against the ECMWF-ENS baseline and about 11 knots in intensity against HWRF.

The model runs on an input grid of roughly 28×28 km — a resolution described as about a hundred times coarser than that of traditional regional models — and generates up to 1,000 scenarios per cyclone, against the roughly 50 members of previous operational ensembles, in less than a minute on a single TPU, over a stated 15-day horizon. The authors admit, however, that they have not fully explained how the system extracts the intensity signal from data at such low resolution, flagging the point as an open research question. The official google-deepmind/weathernext repository publishes the code and the pointers to the checkpoints — the weights are distributed through a Google Cloud bucket — under two stated licences: Apache 2.0 for the Colab notebooks and associated code, Creative Commons Attribution 4.0 for the other materials, which corrects the account of some outlets that referred generically to an Apache licence alone.

The system was used experimentally during the 2025 Atlantic hurricane season in collaboration with the National Hurricane Center (NHC); the announcement cites the case of Hurricane Melissa, whose rapid intensification and landfall in Jamaica the model is said to have anticipated. That account, however, comes from Google DeepMind's release, not from an official NHC report we could locate, and no independent assessment of the specific improvement figures has been published by NHC, ECMWF or the Met Office. Forecasts can be consulted on the Weather Lab platform, while the WeatherNext 2-mini variant can be run from a Colab notebook.

— Olya There is something paradoxical in how the “black box” of neural networks can deliver immediate practical benefits — an extra day to evacuate an area — while science still struggles to explain *why* it works so well on coarse data. Google releasing the weights is a step in the right direction for verification, but until the numbers are validated externally, without the press-release filter, that cited decade of progress remains a commercial promise as much as a scientific achievement.