Overview
- Google says WeatherNext 2 produces hundreds of scenario-based predictions in under a minute on a single TPU at up to hourly resolution.
- The model injects noise through a Functional Generative Network and is trained on marginals to learn spatially coherent joints.
- Google reports performance gains over its prior system on 99.9% of variables and 0–15 day lead times.
- Forecasts now power Search, Gemini and Pixel Weather, with Google Maps integration planned in the coming weeks.
- Data and access are available via Earth Engine and BigQuery, with a Vertex AI early-access program for custom inference, and the system runs four 6-hour cycles daily.