Google DeepMind and Google Research released WeatherNext 3 on Thursday, The Verge and TechCrunch reported. The model is already feeding weather results in Search, Maps, Gemini, Google’s weather API, and Earth Engine.
WeatherNext 2 issued a forecast every six hours on a 25-kilometer grid. WeatherNext 3 can refresh hourly and draw some surface variables, including temperature and moisture, at 5 kilometers, The Verge said. Google calls that a global picture about five times sharper.
The new trick is live satellite mosaics. Most AI weather models train on the delayed analysis that physics supercomputers produce. WeatherNext 3 also ingests raw geostationary satellite observations each hour, Google said. That is meant to catch rain and snow that form fast, and to fill gaps where rain gauges are thin, mostly outside the US and Europe.
Google told The Verge that people planning a day or more ahead will see precipitation forecasts up to 50 percent more accurate. TechCrunch, citing Google’s rain evaluations, said scores improved as much as 60 percent versus WeatherNext 2. Those are company numbers. TechCrunch also reported WeatherNext 3 as the top model on Brightband’s Operational WeatherBench live tests, beating other AI systems and, in that comparison, traditional National Weather Service and ECMWF forecasts.
Google called it the first AI model to pull raw observations into a high-resolution global forecast. WindBorne told TechCrunch its WeatherMesh 6 has used raw balloon and other observations since late 2025. Google’s reply was that its grid is finer worldwide. TechCrunch noted both still lean on national weather datasets.
The Verge said WeatherNext 3 is still trained on physics-based model data and is meant to sit beside those models, not replace the local weather service. Google also added 100-meter wind speeds and solar radiation fields for renewable operators.
The forecast in your Maps pin got a new engine. The warning on your phone still belongs to the National Weather Service.