Google has released version 3 of its WeatherNext AI weather forecasting model, with the key upgrade being its ability to ingest raw satellite data rather than relying solely on "reanalysis" datasets. This reduces the lag between real-world conditions and forecast generation, allowing the model to update hourly instead of every six hours, and matters because AI weather models already offer forecasting performance close to traditional physics-based systems while needing far less computing power, making more frequent updates highly valuable.

Beyond the satellite data, WeatherNext 3 has higher spatial resolution, a larger machine-learning model, and a separate satellite-trained precipitation forecasting component. It also incorporates limited physical information, such as land/ocean status and surface elevation, to improve local temperature and dew point predictions. Google's white paper reports roughly a 5% improvement in upper-atmosphere accuracy over WeatherNext 2 (equivalent to about six extra hours of reliable forecasting) and up to a 30% improvement in surface temperature accuracy, generally outperforming the European Centre for Medium-Range Weather Forecasts' AI model, though some unexplained quirks remain, including weaker early-forecast accuracy for certain variables and visible grid-shaped artefacts in precipitation maps.