DeepMind open-sources cyclone model, adds a warning day

DeepMind open-sourced WeatherNext Cyclones on August 6, per the Nature paper. The model matches three-day track, intensity, and wind-structure accuracy that was previously achievable only two days out.

That extra day matters most when a storm strengthens faster than forecasters expect. WeatherNext Cyclones catches rapid intensification by sampling 1,000 possible forecast tracks instead of committing to one, a technique DeepMind calls Functional Generative Networks.

The model trained on 20 terabytes of weather data plus IBTrACS. DeepMind released the code and weights on GitHub in three variants under an Apache 2.0 license. Any weather service can run, modify, or redistribute it without paying DeepMind. One variant, WeatherNext 2-mini, runs in a free Colab notebook, no in-house supercomputer required.

The National Hurricane Center already used the model operationally for Hurricane Melissa's landfall in Jamaica, per DeepMind. The Nature paper doesn't say whether any other national weather service has started running it since the GitHub release.