Google DeepMind's WeatherNext 3 AI Model Delivers 5km Global Forecasts Refreshed Every Hour

WeatherNext 3 is a Functional Generative Network mesh transformer, the same probabilistic family introduced with WeatherNext 2, scaled to multi-resolution output. It addresses two problems that have limited earlier AI models: resolution too coarse for local terrain, and initialization tied to numerical weather prediction analysis that typically arrives about six hours late. The new model takes the satellite mosaic as a direct input and re-initializes every hour. Its 0.05-degree temperature and dew-point outputs are calibrated against raw station measurements rather than only against reanalysis grids, which often smooth away local variations produced by coastlines, valleys and mountains.

A single forward pass produces three tiers of output: station-trained 2-meter temperature and dew point at roughly 5 km; gridded surface wind at 10 meters and 100 meters, pressure, sea surface temperature, cloud layers, solar radiation and 1-hour precipitation at about 10 km; and atmospheric fields across 13 pressure levels at about 25 km. The report notes that WeatherNext 2 produced 0.25-degree fields in 6-hour increments, making the new system roughly five times sharper in spatial resolution.

The model initializes 24 times a day. The 00, 06, 12 and 18 UTC synoptic cycles run out to 15 days, or 360 hours, with 64 ensemble members, while interim hourly runs cover 48 hours. For fast-developing convection, an hourly refresh grounded in current satellite observations differs meaningfully from a six-hourly cycle anchored to lagged analysis, the report explains.

Precipitation forecasting has historically been a weak point for global models, often producing blurred fields that miss storm boundaries. WeatherNext 3 trains against ECMWF reanalysis, NASA's IMERG satellite retrievals and Google's own satellite-radar precipitation reanalysis. Google reports CRPS improvements of up to 60 percent against IMERG, 30 percent against MRMS and 10 percent against rain gauges at early lead times; the research separately states up to a 50 percent reduction in Brier score and CRPS versus NWP baselines when evaluated against IMERG.

For renewable-energy applications, the model outputs 100-meter wind speed near typical turbine hub height, full low/medium/high cloud distributions and both solar irradiance components, SSRD and FDIR. Those variables are useful for grid operators forecasting wind and solar output against demand, according to MarkTechPost.

Access to WeatherNext 3 is partial. Forecast data is available now through BigQuery, Earth Engine and Cloud Storage after an allowlist request, but the model weights are not open source, and on-demand custom inference still runs WeatherNext 2. The paper and technical details are available, the report said.