Google Research and DeepMind have released WeatherNext 3, a forecasting model that starts from real-time geostationary satellite observations rather than the numerical weather simulations used to train many earlier AI systems. Google says this removes a roughly six-hour delay that can matter for rapidly changing rain and temperature.
The model produces a new forecast every hour. Temperature and humidity can be mapped on a five-kilometer grid, other surface variables at ten kilometers, and atmospheric measurements such as wind at 25 kilometers. Its finest output is about five times sharper than WeatherNext 2’s 25-kilometer, six-hour format, allowing terrain and local weather patterns to appear with more detail.
Google also reports precipitation forecasts up to 50 percent more accurate and has added outputs tailored to renewable-energy operations. Forecasts from the system are now used in Google Search, Maps and Gemini, moving the research directly into consumer products.
A learned forecast is not the same as a physical simulation, and company-reported accuracy can vary by region, lead time and weather type. Emergency planners and energy operators will still need calibrated uncertainty, independent evaluation and conventional observations. The immediate advantage is faster assimilation of current satellite data, not the elimination of meteorological expertise.