Google's new AI weather model doesn't just refine forecasts — it rebuilds them from live satellite data every hour.
WeatherNext 3, built by Google DeepMind and Google Research, hooks into real-time global satellite observations to regenerate global forecasts every hour. That's a serious upgrade over WeatherNext 2, which worked on a 25-kilometer grid with six-hour time steps. The new model resolves surface variables like temperature and humidity down to around 5 kilometers.
Rain and snow are a key focus. Google trained WeatherNext 3 on precipitation data from NASA satellites and radar data. The company says that for forecasts a day or more out, precipitation accuracy can improve by up to 50 percent. Older AI weather models often turned localized storms into vague, blurry blobs; this model is built to capture the sharp details of fast-moving weather.
WeatherNext 3 is already rolling out across Google's ecosystem, including Search, Gemini, Maps, the Google Maps Platform Weather API, and Earth Engine. Developers can also pull forecast data via BigQuery, Earth Engine, and Google Cloud Storage.
Traditional physics-based weather models aren't going anywhere yet. WeatherNext 3 was still trained using output from those models, and meteorological agencies will keep combining multiple forecast systems when issuing disaster warnings. As Google itself cautions, for severe weather and public safety alerts, you should defer to local weather authorities.