Google DeepMind and Google Research have launched WeatherNext 3, a model that ingests live geostationary satellite mosaics and historical observations into a single Functional Generative Network (FGN) mesh transformer to produce hourly, high‑resolution forecasts. Surface temperature and moisture are output on a 5 km grid, other surface variables on a 10 km grid, and atmospheric wind fields on a 25 km grid, delivering roughly five times the spatial detail of WeatherNext 2.
Unlike traditional numerical weather prediction (NWP) which suffers a six‑hour data lag, WeatherNext 3 learns directly from raw satellite observations, enabling rapid capture of fast‑changing phenomena such as thunderstorms and fronts. Precipitation is trained on NASA’s IMERG and an internal radar‑based reanalysis; medium‑range CRPS improvements reach up to 60% over baseline datasets.
The model also predicts 100‑meter wind speeds, high‑resolution cloud cover, and ground‑level solar radiation to support renewable‑energy output forecasting. Forecasts are now available through Google Search, Maps, Gemini, the Google Maps Platform Weather API, BigQuery, and Earth Engine, allowing developers to query or bulk‑download global hourly grids without running their own model.
For regions historically underserved by high‑resolution forecasting—Latin America, Africa, and the Asia‑Pacific—WeatherNext 3’s 5 km grid provides unprecedented local detail, benefiting agriculture, supply chains, and emergency responders. As observation streams grow, the model is poised to further narrow forecast errors, bringing predictions ever closer to on‑the‑ground reality.
Review: WeatherNext 3’s fusion of real‑time satellite data with a large‑scale generative network sets a new benchmark for spatiotemporal resolution and precipitation accuracy, delivering a powerful AI‑driven weather tool for energy, farming, and public safety.