Google DeepMind has published peer-reviewed results in Nature for WeatherNext Cyclones, an AI weather model that achieves state-of-the-art accuracy in predicting tropical cyclone track, intensity, and wind structure. On cyclones from 2023–2025, the model delivers an average of one full day's lead-time advantage over leading operational models (ECMWF-ENS, HWRF)—meaning three-day forecasts rival what prior systems provided at two days. The model runs a 1,000-member ensemble on a single TPU in under a minute, generating probabilistic hazard maps for rare intensification events. Google is now open-sourcing three variants: WeatherNext Cyclones (the version that ran during the 2025 hurricane season), WeatherNext 2, and WeatherNext 2-mini (compact version for free Colab).
The model unifies a longstanding forecasting dilemma: global models excel at track (where the cyclone goes) but fail on intensity (how strong it gets), while local high-resolution models do the opposite. WeatherNext bridges both by learning from a dual-modality dataset: 20 terabytes of global atmospheric reanalysis plus the 5,000-storm IBTrACS historical cyclone database. Unlike physics-based models, it requires only 28x28km input resolution (100x coarser than traditional regional models), achieving this through Functional Generative Networks that efficiently capture aleatoric uncertainty. In operational use during 2025, it helped the U.S. National Hurricane Center issue early warnings for Hurricane Melissa's rapid intensification in Jamaica.
For meteorologists, emergency managers, and climate-dependent sectors (agriculture, insurance, renewable energy), the open release matters more than the Nature paper: a decade's meteorological progress compressed into free, downloadable weights. The one-day lead time corresponds to critical evacuation and resource-staging decisions. Forecasters can now explore 1,000 alternative scenarios per cyclone, shifting prediction from deterministic to probabilistic risk. The practical impact: communities gain ~24h more time to prepare for life-threatening impacts, a margin that meteorologists describe as transformative. For AI practitioners, the result reveals that high resolution is not a prerequisite for extreme-weather modeling when trained end-to-end on large specialized datasets.