DeepMind releases WeatherNext 3

Google DeepMind and Google Research introduced WeatherNext 3, the third generation of their AI weather model, which generates forecasts every hour rather than the six-hour steps of its predecessor and resolves surface variables down to 5km, roughly five times sharper than WeatherNext 2’s 25km grid. The model trains directly on live geostationary satellite data, sparse weather-station readings and satellite precipitation records, rather than relying solely on the outputs of conventional numerical weather prediction systems.

Google reported precipitation-accuracy gains against independent satellite and radar benchmarks, and said independent live evaluations by Brightband ranked it among the most accurate forecasting systems tracked. The model now feeds forecasts into Google Search, the Gemini app, Google Maps and the Maps Platform Weather API, with bulk and query access for developers through BigQuery, Earth Engine and Cloud Storage.

The release is an incremental step in a series rather than a change of approach — Google has now shipped three WeatherNext generations in successive years — but the shift to hourly, satellite-native forecasting extends AI weather models further into short-range and nowcasting use cases that traditional physics-based forecasting handles less well.