Google DeepMind released WeatherNext 3 on September 3, 2026, calling it the most accurate global weather model the Google DeepMind lab has built to date, based on independent live evaluations by the forecasting benchmark group Brightband. The system is already live, powering forecasts in Google Search, the Gemini app, Google Maps and the Maps Platform Weather API.
What changed
The jump in accuracy comes from a change in what the model learns from. Its predecessor, WeatherNext 2, updated every six hours on a 25-kilometer grid, largely mirroring the update cycle of the numerical weather prediction systems it was trained on. WeatherNext 3 instead trains directly on hourly geostationary satellite imagery, NASA’s IMERG satellite precipitation data and ground station readings, which Google says removes the six-hour lag baked into conventional forecasting. The result: forecasts refresh every hour, and surface conditions resolve down to five kilometers, roughly five times sharper than before. Google reports precipitation forecasts up to 50% more accurate a day or more ahead, with early lead-time skill scores improving as much as 60% against satellite rainfall measurements.
New for renewable energy
The model also adds forecasts built for renewable energy planning: wind speed at 100 meters — about turbine height — along with cloud cover and solar radiation, which grid operators can use to estimate wind and solar output before it happens.
WeatherNext 3 extends a broader push by DeepMind to apply machine learning to scientific problems where physics-based simulation has traditionally dominated, an approach the lab groups under its work in AI for science. Beyond its consumer integrations, Google is opening the model’s output to outside researchers through BigQuery, Earth Engine and Cloud Storage — extending its reach into climate and agricultural forecasting work well beyond Google’s own products.