DeepMind open-sources WeatherNext, an AI model that beats top systems at cyclone forecasts

Google DeepMind said on August 6, 2026 that its WeatherNext AI model forecasts tropical cyclone track, intensity and wind structure more accurately than leading physics-based systems. That accuracy gives forecasters roughly an extra day of lead time. The lab open-sourced the code and model weights on GitHub, releasing the code under the permissive Apache 2.0 license.

An extra day of warning on a hurricane’s path and strength can change evacuation orders. DeepMind said the model beats the European ECMWF ensemble on track forecasting and the US HWRF model on intensity — the two systems operational forecasters lean on.

WeatherNext generates a 1,000-member ensemble forecast in under a minute on a single tensor processing unit, at a 28-by-28-kilometer resolution roughly 100 times coarser than traditional models. It was trained on nearly 20 terabytes of atmospheric data plus about 5,000 historical storms from the IBTrACS database, and was co-developed with Google Research, NOAA’s National Hurricane Center, the Cooperative Institute for Research in the Atmosphere and the UK Met Office.

The model has already run in the field. NOAA’s National Hurricane Center used it during the 2025 hurricane season to help forecast Hurricane Melissa’s rapid intensification and Jamaica landfall.

The accuracy claims are DeepMind’s own and rest on retrospective benchmarks rather than a peer-reviewed head-to-head. The coarse resolution limits fine-grained detail near landfall, and physics-based systems still anchor official forecasts.

Open weights let national weather agencies and outside researchers run and scrutinize the model themselves, rather than take the benchmark numbers on faith.