Overview
- Google DeepMind and Google Research announced WeatherNext 3 on Thursday, September 3, and began rolling its forecasts into Search, Maps, Gemini and Google Cloud services.
- WeatherNext 3 generates new hourly global forecasts with native grid detail down to about 5 kilometers for key surface variables and 64‑member probabilistic ensembles with select 15‑day cycles.
- Google reports large precipitation gains, saying day‑ahead rain and snow forecasts can be up to 50% more accurate and showing CRPS improvements in comparisons versus IMERG, MRMS and gauge‑based tests.
- The model trains on raw hourly geostationary satellite mosaics plus sparse station observations to cut the six‑hour latency typical of numerical weather prediction and to better capture fast, local storms.
- Independent live benchmarks on Brightband’s Operational WeatherBench rank WeatherNext 3 ahead of leading AI and some traditional models, while agencies and businesses still need to validate, integrate and govern its operational use.