Google releases WeatherNext 3 AI weather model

Images

×

Image description the-decoder.com

WeatherNext 3 uses a five-kilometer grid for surface temperature, picking up mountain ranges and coastlines that blur into blocky shapes in WeatherNext 2's 25-kilometer grid. | Image: Google Google

For precipitation over the Pacific Northwest, WeatherNext 3 resolves narrow rain bands much more sharply than its predecessor and comes closer to radar observations. | Image: Google Google

The model ingests geostationary satellite data and traditional analysis data, then produces dense gridded fields for atmosphere and precipitation plus point forecasts for weather stations and cyclone tracks. | Image: Google Google

‹ ›

Google and DeepMind have released WeatherNext 3, an AI forecasting system that updates hourly from live geostationary satellite observations rather than running traditional physics-based simulations, according to The Decoder. The model produces forecasts on a five-kilometer grid and is already being used across Google Search, Maps, and Gemini.

The shift is less about replacing meteorology than replacing the slowest part of it: numerical weather prediction runs that depend on supercomputers and arrive with a built-in lag. The Decoder reports that WeatherNext 3’s predecessors learned from those simulation outputs; WeatherNext 3 instead ingests real-time satellite data and generates a fresh forecast every hour, aiming to reduce the compounding errors that show up when rainfall and temperature change faster than the model refresh cycle. Google says the system delivers precipitation forecasts up to 50% more accurate than earlier versions, and it publishes different variables at different resolutions — with temperature and humidity at five kilometers while some atmospheric values are coarser.

The practical winners are not just consumers checking a phone widget. WeatherNext 3 is described as producing data tailored to renewable-energy operations: wind speeds at around turbine height, cloud cover, and solar irradiance, all of which feed into the day-to-day problem of balancing supply and demand. Google is also making the data queryable through BigQuery and Earth Engine, and available for bulk download via Google Cloud Storage, turning what used to be a national-meteorological-service workflow into a cloud product that can be embedded in trading desks, grid-control rooms, and logistics dashboards.

That productization also changes accountability. A forecast based on a physics model can be interrogated in terms of equations and boundary conditions; a forecast learned directly from observation streams is judged by error metrics and backtests. The Decoder notes that regions in Latin America, Africa, and parts of the Asia-Pacific may benefit most, partly because high-compute regional models are expensive and some areas have been underserved. But a finer grid and faster updates also raise expectations: when a forecast is refreshed every hour, the question becomes less “was the model wrong?” and more “which update did you act on, and why?”

WeatherNext 3’s output is now piped into the same Google surfaces where people increasingly treat AI responses as operational advice rather than reference material. The model updates every hour; the decisions it informs can update even faster.