Overview
- WeatherNext 2 generates hundreds of global scenarios from a single starting point in under a minute on one TPU, delivering hourly forecasts up to 15 days ahead.
- The model is already powering weather in Search, Gemini, Pixel Weather, and the Maps Weather API, with broader Google Maps integration coming in the weeks ahead.
- A new Functional Generative Network injects targeted randomness to produce physically coherent ensembles, learning from single-variable ‘marginals’ to predict multi-variable ‘joints’.
- Google reports the system surpasses its previous model on 99.9% of variables and lead times, with published CRPS gains over GenCast indicating stronger probabilistic skill.
- DeepMind cites improved tropical cyclone tracking and says experimental cyclone tools have been shared with agencies, while acknowledging weaknesses for rare rain and snow events.