Google debuts WeatherNext 3 satellite weather model

Google Research and DeepMind have introduced WeatherNext 3, a weather forecasting model that abandons traditional numerical weather prediction simulations. Instead of relying on physics-based supercomputer simulations that suffer from a six-hour delay, the new system processes real-time geostationary satellite data to generate fresh forecasts every hour. The model produces temperature and humidity forecasts on a five-kilometer grid, which is five times sharper than the 25-kilometer grid used by its predecessor, WeatherNext 2.

To capture regional nuances like coastlines and valleys, WeatherNext 3 trains on individual weather station data alongside NASA's satellite-based IMERG dataset and Google's own global precipitation analysis. The model provides surface variables on a 10-kilometer grid and atmospheric values like wind speed on a 25-kilometer grid. According to Google, these updates yield precipitation forecasts that are up to 50 percent more accurate. For short lead times, the model's Continuous Ranked Probability Score shows improvements of up to 60 percent over IMERG, 30 percent over MRMS, and 10 percent over traditional rain gauges.

For industry practitioners, this model introduces critical tools for the renewable energy sector. WeatherNext 3 estimates wind farm output by predicting wind speeds at 100 meters, which is roughly turbine height, while also calculating cloud cover and solar irradiance to help solar installations project energy generation. Developers and researchers can access this data hourly via BigQuery, Earth Engine, or Google Cloud Storage. The technology is already integrated into Google Search, Maps, and the Gemini app.

This release builds on Google's previous meteorological AI projects. WeatherNext 2, launched in November 2025, was built on the same Functional Generative Network architecture but relied on older training methods. DeepMind also open-sourced WeatherNext 2 and WeatherNext Cyclones in August 2026, following the late 2024 release of GenCast, its first probabilistic weather model.