DeepMind’s WeatherNext 3 adds hourly satellite-fed forecasts

Google DeepMind and Google Research published WeatherNext 3, a global weather model they say takes live weather-satellite mosaics and refreshes hourly at about 5 kilometers on key surface fields.

If you use Google Search or Maps, the weather there may now come from an hourly satellite-fed model. The 5 kilometer grid and the 50% precipitation line stay DeepMind’s until someone else measures them.

DeepMind posted “Introducing WeatherNext 3.” The model, built with Google Research, takes live geostationary satellite mosaics — Earth-facing weather-satellite composites — as a direct input and issues a new forecast every hour.

DeepMind says key surface fields such as temperature and moisture are shown at about 5 kilometers, other surface fields at 10 kilometers, and atmospheric fields at 25 kilometers. It calls that roughly five times sharper than WeatherNext 2, which ran on a 25 kilometer grid in six-hour steps.

The lab cites independent live evaluations by Brightband’s Operational WeatherBench. It also says longer-range precipitation in Search, Gemini, Maps, the Maps Platform Weather API, and Earth Engine is up to about 50% more accurate. Those figures are DeepMind’s.

Hourly data is offered through BigQuery, Earth Engine, and Cloud Storage downloads. The post says the model is already wired into Search, Gemini, Maps, Maps Platform, and Cloud.

DeepMind’s own disclaimer: official forecasts and severe-weather warnings still belong to national weather services. The weights are not described as public.