DeepMind open-sources weather AI that buys forecasters an extra day

Google DeepMind open-sourced WeatherNext, an AI model that gives storm forecasters an extra 24 hours of advance warning.

In a paper published in Nature, researchers detailed WeatherNext Cyclones, a system that predicts cyclone tracks, intensity, and wind structure. Its three-day forecast matches the accuracy of older two-day forecasts, matching roughly ten years of typical progress in meteorology. Instead of relying on traditional supercomputer setups, it runs a 15-day forecast ensemble in under a minute on a single TPU.

Why it matters: Tropical cyclones caused $1.4 trillion in damage and 700,000 deaths over the last 50 years. An extra day of lead time gives emergency teams vital time to prepare. The system already proved its worth during Hurricane Melissa in 2025, helping forecasters predict its rapid intensification before hitting Jamaica.

Know this: The model operates on coarse 28x28km data — 100 times lower resolution than traditional physics models. DeepMind released the weights and code for WeatherNext 2 and WeatherNext Cyclones so the community can build on it.

Coastal planners just got a massive upgrade, even if the mechanics behind its low-res accuracy are still keeping researchers up at night.