Google DeepMind Launches WeatherNext 3 With 5km Hourly AI Forecasts

Google DeepMind and Google Research released WeatherNext 3. The AI weather model bypasses traditional delayed numerical weather prediction by ingesting live global geostationary satellite mosaics as direct inputs. The system trains dedicated observational heads directly on raw station measurements. This allows it to output 5 km (0.05°) resolution forecasts calibrated to actual instrument recordings rather than smoothed atmospheric models. WeatherNext 3 re-initializes 24 times a day. The model generates 15-day forecasts with 64 ensemble members during major synoptic cycles and handles 48-hour interim runs every hour. The architecture achieves up to a 60% CRPS improvement against NASA IMERG baselines for precipitation. Google also reported up to a 50% reduction in Brier score and CRPS versus NWP baselines. Forecast data is currently gated behind allowlists on BigQuery, Earth Engine, and Cloud Storage. The underlying model weights remain closed source, and on-demand custom inference still relies on the older WeatherNext 2.