Google DeepMind and Google Research have launched WeatherNext 3, a forecasting model that generates hourly predictions at up to 5-kilometer resolution, roughly five times sharper than the 25-kilometer grid its predecessor used in six-hour increments. The system ingests live geostationary satellite imagery instead of relying purely on physics simulations, and it refreshes its picture of the atmosphere every hour.
The model extends the Functional Generative Network approach behind WeatherNext 2, running 15-day forecasts with 64-member ensembles trained to minimize a standard measure of probabilistic forecast skill. The production version was trained on data through June 30, 2026.
Independent live evaluations by the startup Brightband on its Operational WeatherBench leaderboard put WeatherNext 3 ahead of rival deep-learning systems from Microsoft, Nvidia, and the European Center for Medium-Range Weather Forecasting, and above conventional forecasts from the US National Weather Service and ECMWF.
Google is wiring the model into Search, the Gemini app, Google Maps, the Google Maps Platform weather API, and Earth Engine. Senior staff engineer Samier Merchant told TechCrunch this is the first time core weather variables will power so many Google products, and researchers can reach the system through Google Cloud.
The launch lands as AI weather models displace expensive government supercomputer runs, with sharper outputs meant to help agriculture, clean-energy planning, and warnings for heatwaves, storms, and drought.
