Google DeepMind has released WeatherNext 3, an AI weather model that initialises forecasts every hour using live satellite data to deliver global predictions at a 5 km resolution. This approach addresses two persistent issues in AI forecasting: the inability to resolve local terrain features and the reliance on numerical weather prediction analysis that arrives six hours late.
Architecture and inputs
The system is a Functional Generative Network, or FGN, mesh transformer. It draws on the same probabilistic family introduced with WeatherNext 2 but scales to multi-resolution output. Inputs include a live global geostationary satellite mosaic and ECMWF HRES analysis. Training data combines ERA5/HRES-fc0, NASA’s IMERG, station observations, and satellite mosaics.
Most AI forecasters learn from NWP reanalysis, which smooths away local variation caused by coastlines, valleys, and mountains. WeatherNext 3 trains dedicated observational heads directly on raw station measurements. Consequently, its 0.05° temperature and dew point outputs are calibrated to what instruments record rather than a model’s representation of the atmosphere.
Resolution and cadence
A single forward pass produces three tiers of data. The first is 0.05° (~5 km) station-trained 2 m temperature and dew point. The second is 0.1° (~10 km) gridded surface wind at 10 m and 100 m, pressure, sea surface temperature, cloud layers, solar radiation, and 1-hour precipitation. The third tier covers 0.25° (~25 km) atmospheric fields across 13 pressure levels.
WeatherNext 2 produced 0.25° fields in 6-hour increments. The roughly 5x sharper claim for WeatherNext 3 stems from this shift to hourly output. The model initialises 24 times a day. The 00, 06, 12, and 18 UTC synoptic cycles run out to 15 days with 64 ensemble members. Interim hourly runs cover 48 hours. For fast-developing convection, an hourly refresh grounded in current satellite observations is meaningfully different from a six-hourly cycle anchored to lagged analysis.
Precipitation and clean energy variables
Precipitation is where global models historically fail, producing blurred fields that miss storm boundaries. WeatherNext 3 trains against three precipitation sources: ECMWF reanalysis, NASA’s IMERG satellite retrievals, and Google’s own satellite-radar precipitation reanalysis.
Google reports CRPS improvements over baselines of up to 60% against IMERG, 30% against MRMS, and 10% against rain gauges at early lead times. The research separately states up to a 50% reduction in Brier score and CRPS versus NWP baselines when evaluated against IMERG.
For renewables, the model outputs 100 m wind speed at approximate turbine hub height, full low/medium/high cloud distributions, and both solar irradiance components (SSRD and FDIR). That combination is what grid operators need to forecast wind and solar output against demand, and it is the clearest sign that this release is aimed at operational buyers, not only at benchmark tables.
Deployment status
Forecast data is available now through BigQuery, Earth Engine, and Cloud Storage after an allowlist request. WeatherNext 3 weights are not open source, and on-demand custom inference still runs WeatherNext 2.
What it means
Operational users gain access to hourly updates without waiting for the standard six-hour analysis lag. The finer resolution allows for better prediction of localised weather events, which is critical for grid stability and energy planning.
