Google’s Weather AI Gets Further Speed Boost: 5km-Resolution Forecasts, Updated Hourly

Beating AI News Flash: Google DeepMind and Google Research have released WeatherNext 3, a global weather forecasting model integrated directly with real-time global satellite observations, capable of regenerating global forecasts every hour. The previous WeatherNext 2 operated on a 25-kilometer grid with 6-hour intervals; the new model boosts maximum grid resolution to approximately 5 kilometers for surface variables like temperature and humidity. Precipitation and snowfall forecasting were key focus areas for improvements. WeatherNext 3 was trained using NASA satellite precipitation data and radar data. Google says that for forecasts of one day or longer, precipitation prediction accuracy can improve by up to 50%. Traditional AI weather models often produced overly vague forecasts for localized heavy rainstorms; this update primarily enhances details for such rapidly changing weather events. WeatherNext 3 is now integrated with Google Search, Gemini, Maps, the Google Maps Platform Weather API, and Earth Engine. Developers can also access forecast data via BigQuery, Earth Engine, and Google Cloud Storage. Traditional physical weather models will not disappear anytime soon. WeatherNext 3’s training still incorporates data from these models, and meteorological agencies typically combine multiple forecast sets when issuing disaster warnings. Google itself reminds users that severe weather and public safety warnings should still be based on local meteorological authorities.