Google DeepMind open sources WeatherNext Cyclones and the WeatherNext 2 models alongside a Nature paper

Sources: primary, discussion
Summary: Google DeepMind's post of 2026-08-06 open sources three models: WeatherNext Cyclones, WeatherNext 2, and WeatherNext 2-mini. DeepMind states that WeatherNext Cyclones gains more than a full day of lead time on cyclone track, intensity and wind structure, so its three-day forecasts match what prior models gave at two days, which DeepMind puts at roughly a decade of meteorological progress. It reports that result at 28x28km resolution, 100x coarser than traditional models. The accuracy claim is published in a Nature paper. WeatherNext 2-mini is the compact variant that runs on a single TPU, and it is part of the WeatherNext 2 release rather than the cyclone models.
Why it matters: Code and weights for a model whose accuracy claim went through peer review give teams a reusable forecasting component instead of a vendor API.