Google DeepMind launches WeatherNext 3 with 5 km global forecasts and hourly refresh

The system builds on the same Functional Generative Network mesh transformer architecture introduced in version 2, now calibrated for multi-resolution output. Its primary input is a live geostationary satellite mosaic rather than a delayed numerical weather prediction analysis that arrives six hours late, supplemented by ECMWF HRES analysis. Training drew on ERA5, HRES-fc0, NASA IMERG, ground-station observations and satellite mosaics. The fundamental difference: dedicated forecast heads were trained directly on raw station measurements, so the 0.05° temperature and dew-point output is calibrated to what instruments actually record, not to a physical model's atmospheric representation.

A single forward pass produces three layers. At 0.05° (roughly 5 km) it delivers station variables, 2-metre temperature and dew point. At 0.1° (roughly 10 km) it provides surface grid fields — 10-metre and 100-metre wind, pressure, sea-surface temperature, cloud layers, solar radiation and hourly precipitation. At 0.25° (roughly 25 km) it covers atmospheric fields across 13 pressure levels. WeatherNext 2 supplied 0.25° fields at six-hour intervals, so the claimed five-fold sharpening comes from direct comparison. The cadence is the second change: initialisation runs 24 times a day. Synoptic cycles at 00, 06, 12 and 18 UTC extend to 15 days (360 hours) with 64 ensemble members; hourly interim runs cover 48 hours. For fast-evolving convection, an hourly refresh anchored in current satellite observations is fundamentally different from a six-hour cycle anchored in lagged analysis.

Precipitation has been the historical failure point of global models, which produce blurry fields and miss storm boundaries. WeatherNext 3 was trained against three precipitation sources: ECMWF reanalysis, NASA IMERG satellite retrievals, and Google's own satellite-radar precipitation reanalysis. According to Google, CRPS improvement versus baselines reaches up to 60% against IMERG, 30% against MRMS and 10% against rain gauges at early lead times; the separate research note cites up to 50% reduction in Brier score and CRPS versus NWP baselines when evaluated against IMERG. For renewables the model outputs 100-metre wind speed — typical turbine hub height — low, mid and high cloud fractions, and both solar radiation components (SSRD and FDIR). That combination is exactly what grid operators need to forecast wind and solar output against demand, and it is the clearest signal that the release targets operational buyers, not just benchmark tables.

Forecast data is now available through BigQuery, Earth Engine and Cloud Storage after an allowlist request, but the WeatherNext 3 weights are not open source and on-demand adapted inference still runs on WeatherNext 2. An independent live evaluation by Brightband ranks it the most accurate global weather model to date, yet deployment remains partial: without open weights and without an adapted inference path in the new version, the ability to run the model end-to-end in your own environment is still blocked.