Google Unveils WeatherNext 3 AI Forecasting Model

Weather forecasting has traditionally relied on government agencies running physics-based numerical models on costly supercomputers. Deep-learning systems can generate comparable predictions faster, but have often struggled with local detail, rainfall and delays in assimilating observations. Developed by Google DeepMind and Google Research, WeatherNext 3 is designed to learn directly from live satellite and weather-station data, a shift that could improve decisions in disaster response, agriculture, transport and renewable-energy operations.

Google unveiled WeatherNext 3 on Sept. 3, 2026, with hourly global forecasts and resolution as fine as 5 kilometers for key surface variables, versus WeatherNext 2’s 25-kilometer grid and six-hour intervals. Google reported precipitation CRPS gains of as much as 60% against IMERG-based benchmarks and up to 50% better precipitation forecasts more than a day ahead. Brightband’s Operational WeatherBench ranked it ahead of models from Microsoft, Nvidia, ECMWF and the U.S. National Weather Service. The system began powering Search, Gemini, Maps and Google cloud products the same day.