Google DeepMind and Google Research have announced the introduction of WeatherNext 2, their latest and most efficient weather forecasting model. This new model completes calculations on a single TPU in less than one minute—tasks that previously required several hours on supercomputers using traditional physics-based models. Processing speed has improved eightfold.
Prediction resolution can be refined down to one-hour intervals. The model generates hundreds of scenarios from a single initial state, capturing a wide range of possibilities, including worst-case scenarios. Compared to the previous generation model, WeatherNext, it outperforms in 99.9% of variables, such as temperature and wind speed, across forecast periods from 0 to 15 days.
On the technical side, the model adopts a new approach called Feature Generation Networks (FGN). By injecting noise into the model architecture, it ensures predictions that are physically realistic and mutually consistent. While training on individual weather elements, it also improves the prediction accuracy of the complex overall system in which those elements combine.
Prediction data from WeatherNext 2 is currently accessible via Earth Engine and BigQuery. Additionally, an early access program for custom model inference has been launched on the Google Cloud Vertex AI platform. This technology is also being integrated into Search, Gemini, Pixel Weather, and the Google Maps Platform Weather API, and will be reflected in Google Maps weather information in the near future.
Source: WeatherNext 2: Our most advanced weather forecasting model (HN 291pt, 131 comments) (HN Search (backfill), 2025-11-18)