Google DeepMind just built a weather forecasting model that outperforms the humans who’ve spent their careers tracking hurricanes. WeatherNext 2, developed in collaboration with Google Research, can predict both the path and intensity of tropical cyclones with accuracy that surpassed traditional physics-based methods during the 2025 Atlantic hurricane season.
What WeatherNext 2 actually does
The model generates probabilistic ensemble forecasts up to 15 days ahead, with temporal resolution down to one hour. It runs 8x faster than its predecessor models. The National Hurricane Center tested WeatherNext 2 during the 2025 hurricane season. It outperformed conventional methods across several key performance metrics for cyclone forecasting. Perhaps most intriguing, the researchers behind the model don’t yet fully understand how it achieves this level of accuracy using lower-resolution weather data inputs.
Earlier components of the WeatherNext series, along with related models like GraphCast and GenCast, have already been open-sourced. Pretrained weights and lower-resolution data inputs are publicly available through Google Developer resources. DeepMind has indicated that WeatherNext itself will also be open-sourced, which means anyone with the technical chops can build on top of it.
