On September 3, 2026, Google DeepMind and Google Research introduced WeatherNext 3, the company's most advanced global weather AI model to date. The system ingests live geostationary satellite mosaics every hour and produces forecasts at up to 5-kilometer resolution for surface temperature and dew point, 10-kilometer resolution for other surface variables, and 25-kilometer resolution for atmospheric pressure levels. The model is now integrated across Search, Gemini, Maps, Google Maps Platform, and Cloud.

WeatherNext 3 replaces the six-hour update cadence of its predecessor with hourly initialization — 24 cycles per day — and expands the ensemble to 64 members. The architecture is a Functional Generative Network (FGN) mesh transformer that consumes raw satellite imagery alongside ECMWF HRES analysis, then outputs dense gridded fields, discrete cyclone tracks, and station-level sparse coordinates in a single forward pass.

Confirmed

  • WeatherNext 3 was announced by Google DeepMind and Google Research on September 3, 2026, with hourly satellite-driven initialization, 64-member ensembles, and the multi-resolution output grid described above.
  • The model is available via BigQuery, Earth Engine, and Google Cloud Storage (Zarr), and is integrated into Search, Gemini, Maps, Google Maps Platform, and Cloud.
  • Google reports up to 60% CRPS improvement versus IMERG, 30% versus MRMS, and 10% versus rain gauges for early lead times — vendor-reported figures, not independently verified.
  • Google says station-trained 5 km heads match ground observations more closely than reanalysis-grid training alone — again a company claim.

What's new

  • Hourly refresh: Live geostationary satellite mosaics feed the model every hour, cutting the data lag that limited earlier AI weather models to six-hour cycles.
  • Multi-resolution output: 0.05° (~5 km) for station-trained 2 m temperature and dew point; 0.1° (~10 km) for gridded surface wind (10 m and 100 m), pressure, sea-surface temperature, cloud layers, solar radiation, and precipitation; 0.25° (~25 km) for 3D atmospheric variables on 13 pressure levels (available in the 6-hourly synoptic cycles).
  • Ground-truth station training: Dedicated observational heads train directly on raw weather-station measurements; Google says this produces 5 km outputs that match ground observations more closely than reanalysis-grid training alone.
  • Precipitation: Trained on NASA IMERG, Google's satellite-radar reanalysis, and ECMWF reanalysis, with the CRPS comparisons Google published for early lead times.
  • Clean-energy variables: 100-meter wind speeds (turbine-height), full cloud-layer distributions, and complete solar irradiance components (SSRD, FDIR) for renewable energy planning.
  • Forecast horizon: 15 days (360 hours) for 6-hourly cycles; 48 hours for interim hourly runs.
  • Access: BigQuery, Earth Engine, and Google Cloud Storage (Zarr). Full ensemble Zarr is exclusive to Google Cloud Storage for 6-hourly cycles.

Why it matters

Hourly AI forecasts that start from live satellite mosaics lower the barrier for high-resolution guidance in regions without local supercomputing — Latin America, Africa, and parts of Asia-Pacific among them. For energy markets, 100-meter wind and full solar-irradiance fields at 10 km give grid operators and developers a denser signal than day-ahead NWP alone. Google also pitches sharper convective precipitation; that is the claim to watch, not a settled operational fact.

Unknown

  • The CRPS improvement figures (60% vs IMERG, 30% vs MRMS, 10% vs rain gauges) are Google's own reported benchmarks; no independent evaluation has yet been published.
  • Google's claim that station-trained 5 km outputs match ground observations more closely than reanalysis-grid training has not been independently validated.
  • Real-world accuracy in operational use — particularly for sharp convective precipitation — remains unproven outside Google's evaluations.

Our take

WeatherNext 3 suggests that end-to-end learning from raw satellite pixels may be closing in on — and in some settings could outperform — hybrid approaches that still lean on numerical weather prediction analyses for initialization; that is an editorial reading of Google's disclosure, not a settled result. The remaining dependency on ECMWF HRES analysis means the pipeline is not yet fully observation-driven. Until assimilation skips that NWP step, the hourly cadence and 5 km station heads are Google's strongest product bet — especially in Search, Maps, and Gemini — more than a proven edge for energy traders.

Sources