Google DeepMind and Google Research have introduced WeatherNext 3, the most advanced global weather AI model to date, featuring real-time satellite data, hourly refreshes, and five times higher resolution than its predecessor. Independent live evaluations by Brightband confirm its superior accuracy. The model generates hourly forecasts at multiple spatial resolutions—5-kilometer resolution for key surface variables like temperature and moisture, 10-kilometer for other surface variables, and 25-kilometer for atmospheric variables like wind speed.

Unlike previous models trained on six-hour lagged numerical weather prediction (NWP) data, WeatherNext 3 learns directly from raw, live geostationary satellite mosaics, enabling more timely and localized predictions. It’s now integrated across Google Search, Gemini, Maps, Google Maps Platform, and Cloud. WeatherNext 3 uses a Functional Generative Network (FGN) mesh transformer to process real-time observations and historical data, producing dense gridded fields, discrete cyclone tracks, and native station-level predictions.

This advancement allows for detailed visualization of local topography and rapid updates critical for tracking sudden weather changes like storms and precipitation systems.