WeatherNext 3 from Google Research and DeepMind drops traditional physics models and learns directly from real-time satellite data. Google promises five times the detail and hourly updates.

Previous AI weather models, including WeatherNext 2, trained on numerical weather prediction (NWP) data. Those simulations run on supercomputers and carry a six-hour delay, according to Google, which introduces errors for fast-changing variables like rainfall and temperature.

WeatherNext 3 processes live geostationary satellite data instead. It generates a fresh forecast every hour based on the latest observations at up to five-kilometer resolution. For fast-developing storms or precipitation, the quicker update cycle means earlier and more accurate predictions.

Five times sharper than its predecessor

The model produces hourly forecasts at multiple resolutions. Temperature and humidity run at five kilometers, other surface variables at ten, and atmospheric values like wind speed at 25. That’s about five times sharper than WeatherNext 2, which used a 25-kilometer grid in six-hour intervals.

Google illustrates the gap with a temperature forecast over the United Kingdom. WeatherNext 2 at 25 kilometers produces a pixelated map. WeatherNext 3 at five kilometers captures local terrain.

The model also trains on individual weather station data to capture coastlines, valleys, and mountains. That enables finer global forecasts that reflect regional conditions. Regions in Latin America, Africa, and the Asia-Pacific stand to gain the most, according to Google, since high compute costs for traditional regional models have left them underserved.

Better rain forecasts and renewable energy data

Global weather models have long struggled with precipitation because rain and snow depend on fast cloud processes at small scales. Google trained the model on two sources: NASA’s satellite-based IMERG dataset and Google’s own global precipitation analysis built on satellite radar.

For medium-range global forecasts, the Continuous Ranked Probability Score (CRPS) shows improvements of up to 60 percent over IMERG, 30 percent over MRMS, and ten percent over rain gauges at short lead times. WeatherNext 3 also predicts wind speeds at 100 meters, roughly turbine height, to estimate wind farm output. More accurate cloud cover and solar irradiance values let solar installations calculate expected generation. Google says this should help grid operators balance supply and demand.

Rolling out across Search, Maps, and Gemini

Google updates the forecast data hourly. Researchers, developers, and businesses can query it through BigQuery and Earth Engine or download it in bulk from Google Cloud Storage. WeatherNext 3 now powers weather features in Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine.

When planning a day or more ahead, users should see up to 50 percent more accurate precipitation forecasts, with the biggest gains in regions where predictions have been less reliable.

Google acknowledges that the atmosphere will always be somewhat unpredictable. For official forecasts, severe weather warnings, and safety advisories, the company points users to national weather services.

Google DeepMind introduced WeatherNext 2 in November 2025. That model beat the first WeatherNext generation on 99.9 percent of all meteorological variables and forecast windows while running eight times faster, processing hundreds of weather scenarios in under a minute on a single TPU. It was built on the Functional Generative Network that also underpins WeatherNext 3. WeatherNext 2 already powered Google Search, Gemini, Pixel Weather, and the Weather API.

In August 2026, DeepMind released WeatherNext 2 and WeatherNext Cyclones under an open license. The cyclone model predicts tropical storm tracks and intensity in a single system and looks about a day further ahead than leading operational models, despite using a grid roughly a hundred times coarser. In late 2024, DeepMind released GenCast, the first probabilistic AI weather model to outperform the ECMWF’s ensemble system.