Google has released the 3rd version of its WeatherNext AI weather model. The update’s most important innovation was the direct integration of some satellite weather data into the model. This reduced the lag between current conditions and the new forecast, while increasing the forecast frequency from once every six hours to hourly.
Most AI weather models rely on “reanalyses” that combine data from different sources into a single global picture of the atmosphere. WeatherNext 3 uses raw satellite data alongside this approach. The model’s spatial resolution was increased, the machine-learning model was expanded, and some processes were modified to limit the computational load. A separate model trained with satellite-based precipitation forecasts was also added, enabling multiple precipitation forecasts to be provided.
When calculating surface temperature and dew point at a specific location, the model takes into account whether the location is on land or at sea, as well as the elevation of the surface. The team says that using this information together with historical weather-station data improves the forecasts.
According to the White paper, WeatherNext 3 improved the accuracy of upper-atmosphere conditions by approximately 5% compared with WeatherNext 2, corresponding to an accurate forecast period approximately six hours longer. For specific locations, it improved the accuracy of surface-temperature calculations by up to 30%. Overall, the model also outperformed the European Centre for Medium-Range Weather Forecasts’ (ECMWF) AI model on these metrics.
However, it was reported to produce worse results in the first six-hour forecasts for some variables, while grid-like hexagonal shapes appeared on some maps. WeatherNext 3 became the source of forecast information in Google Search, Gemini and Maps.
Why it matters
Combining satellite observations with raw data could affect the monitoring of rapidly changing conditions by reducing the time gap between current conditions and model output in weather forecasts. Hourly updates mean that people who use forecast information through Google Search, Gemini and Maps can access more frequently refreshed data. However, the performance gains do not extend across all measurements; the deterioration in the first six-hour results for some variables and the grid-like shapes seen on some maps show that the output cannot be considered more reliable under all conditions. Although the comparison with ECMWF’s AI model was positive, the open question is under which conditions WeatherNext 3’s superiority across different variables is maintained and to what extent the limitations affect the forecasts presented to users.
Background
Google is not a new name in the FikirPilot archive: we have published 28 news reports mentioning the name in the last 90 days; the most recent is dated September 9, 2026.