Three years into the AI weather modeling race, two stubborn problems remained: resolution too coarse for local terrain, and initialization stuck with six-hour-old NWP analysis. Google DeepMind’s WeatherNext 3 takes a swing at both. The model ingests a live geostationary satellite mosaic directly, re-initializes every hour, and pushes output down to 0.05 degrees—roughly 5 kilometers—while training against raw weather station measurements rather than smoothed reanalysis grids.
That last part matters more than it sounds. Most AI forecasters learn from NWP reanalysis, which averages away the local variation that coastlines, valleys, and mountains actually produce. WeatherNext 3 instead trains dedicated observational heads directly on station data, so its temperature and dew point outputs are calibrated to what instruments record. Google reports CRPS improvements of up to 60% against IMERG precipitation data at early lead times, along with 30% against MRMS and 10% against rain gauges. Independent live evaluations from Brightband rank it as the most accurate global weather model to date.
The cadence shift is equally significant. The model initializes 24 times a day—synoptic cycles at 00, 06, 12, and 18 UTC run out to 15 days with 64 ensemble members, while 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. Single forward passes produce three resolution tiers: 5km station-trained variables, 10km gridded surface fields including 100m wind speed and solar irradiance components, and 25km atmospheric fields across 13 pressure levels.
The renewables angle is hard to miss. Outputs include turbine hub-height wind, full cloud distributions, and both solar irradiance components—exactly what grid operators need to forecast supply against demand. That’s the clearest signal this release targets operational buyers, not just benchmark tables. Forecast data is available now through BigQuery, Earth Engine, and Cloud Storage after an allowlist request, but the weights aren’t open source and on-demand custom inference still runs WeatherNext 2.
