Google DeepMind on September 3, 2026 unveiled WeatherNext 3, its most accurate global AI weather model to date – and, the company says, the first to generate a forecast for every hour of the day.
Sharper resolution, hourly updates
WeatherNext 3 delivers temperature and moisture on a 5-kilometre grid, other surface variables at 10 kilometres and atmospheric fields such as wind speed at 25 kilometres. That is roughly five times finer than its predecessor WeatherNext 2, which used a uniform 25-kilometre grid throughout. Rather than running only every six hours like classical models, the system refreshes hourly.
Learning straight from satellite data
Technically, DeepMind relies on a so-called Functional Generative Network, a mesh transformer that ingests live geostationary satellite data alongside historical analysis. Because the model learns directly from raw observations rather than the output of numerical weather models, it removes a six-hour data-refresh lag. For precipitation, Google cites a Continuous Ranked Probability Score improvement of up to 60 percent against the IMERG reference dataset; for medium-range planning a day or more ahead, it claims rainfall forecasts up to 50 percent more accurate.
Value for energy grids and everyday use
Of particular interest to grid operators and renewable-energy developers: WeatherNext 3 predicts wind speeds at 100 metres – roughly turbine hub height – as well as cloud cover and solar radiation, helping estimate output from wind and solar assets. The forecasts already feed Google Search, the Gemini app and Google Maps; researchers and developers can access the data through Earth Engine, BigQuery and Google Cloud Storage.
Sources: blog.google · deepmind.google · qz.com
