Google DeepMind unveils WeatherNext 3, a high-resolution AI weather model with hourly updates

Google DeepMind and Google Research announced WeatherNext 3, a global weather AI model that ingests real-time satellite data to produce hourly forecasts at higher spatial resolution than its predecessor. The model uses 1-hour geostationary satellite mosaics alongside historical analysis to feed a Functional Generative Network (FGN) mesh transformer, outputting dense gridded fields and discrete cyclone tracks.

WeatherNext 3 achieves a global weather picture roughly five times sharper than WeatherNext 2, which previously produced forecasts on a 25-kilometer grid in 6-hour increments. The new model visualizes key surface variables like temperature and moisture at 5-kilometer resolution, other surface variables at 10 kilometers, and atmospheric variables such as wind speed at 25 kilometers.

The model is now integrated across Google products including Search, Maps, Gemini, Maps Platform, and Cloud, enabling users to access enhanced forecasts directly or integrate the data into their own projects.

Independent live evaluations by Brightband are cited as supporting the claim that WeatherNext 3 is the most advanced and accurate global weather model to date.