WeatherNext 3 Directly Incorporates Real-Time Satellite Observations, Upgrading Global AI Weather Forecasting to Hourly Updates and 5-Kilometer Grids

Google DeepMind and Google Research have released WeatherNext 3, advancing global AI weather forecasting from WeatherNext 2’s 25-kilometer resolution and six-hour update cycle to resolutions as fine as 5 kilometers for some variables, with forecasts regenerated every hour. The key change is the incorporation of real-time geostationary satellite imagery, giving the model access to more recent atmospheric observations rather than training exclusively on analysis data produced by traditional numerical weather prediction (NWP), which typically arrives with a delay of about six hours.

This marks an important change in the data pipeline. Rapidly developing precipitation, surface-temperature changes, and cloud systems can easily be smoothed out in low-frequency inputs. WeatherNext 3 uses globally mosaicked satellite data to improve temporal resolution and trains additional outputs for weather-station locations, allowing gridded predictions to be compared directly with ground observations. The model has roughly 2.4 times as many parameters as its predecessor, and its decoder heads are tailored to different targets. Google says precipitation evaluation results improved by about 60% relative to WeatherNext 2, but it does not claim that this translates into a 60% reduction in errors across every region or for all extreme events.

Real-time comparisons from the independent Operational WeatherBench show that WeatherNext 3 leads several AI-based and conventional forecasting systems on metrics including temperature, wind speed, and humidity. However, rankings are affected by forecast initialization times, variables, regions, and validation data, so a single aggregate ranking is insufficient to prove greater reliability for typhoon tracks, short-duration heavy rainfall, or rare extreme events. Google’s claim that WeatherNext 3 is the first system to use raw observations directly for high-resolution global forecasting has also been challenged by WindBorne, which says WeatherMesh 6 already uses balloons and other real-time observations. Both systems still use data from national meteorological agencies, so “direct observations” does not mean that NWP data assimilation has been completely replaced.

WeatherNext 3 is beginning to power Search, Gemini, Maps, Google Maps Platform, and Earth Engine. For engineers working in energy, agriculture, and logistics, the next priorities should be the licensing terms for downloadable outputs, ensemble forecasts and uncertainty representation, regional backtesting, and the latency and service levels of hourly updates in production APIs—not just global average scores.