WeatherNext 3

WeatherNext 3 is the latest model in the WeatherNext family — developed jointly by Google DeepMind and Google Research — for AI-based weather forecasting. It is a generative ensemble model that learns directly from real-time observations (including geostationary satellite-data mosaics) and generates a new forecast every hour.

Unlike earlier models, which were trained mainly on the output of classical numerical weather prediction (NWP) simulations that carry a roughly six-hour data lag, WeatherNext 3 uses a single, flexible Functional Generative Network (FGN) built on a mesh transformer. It delivers forecasts at up to 5 km resolution for temperature and humidity, 10 km for other surface variables, and 25 km for atmospheric variables — roughly 5x sharper than WeatherNext 2, which ran on a 25 km grid in six-hour steps.

The model substantially improves precipitation forecasting (up to about 50% more accurate when planning a day or more ahead) and introduces variables useful for renewable energy: 100-meter wind speed, cloud cover and solar radiation. According to independent live evaluations by Brightband, it is the most accurate global weather model to date.

WeatherNext 3 powers weather experiences in Google Search, the Gemini app, Google Maps and the Google Maps Platform Weather API, while the data is available to developers and businesses through Google Earth Engine, BigQuery and Google Cloud Storage.