DeepMind opens two global weather datasets behind WeatherNext 3

The latest:

Two foundational datasets behind Google DeepMind’s WeatherNext 3 weather model are now open to outside users. In a November 13, 2026 post credited to the WeatherNext team, DeepMind said SatWeave, a global geostationary satellite mosaic, and PARDIG, an AI-generated precipitation record, are available for weather, hydrology and climate applications through Google Cloud Storage, with live feeds alongside historical archives.

Details:

• What SatWeave is: According to DeepMind, SatWeave stitches imagery from GOES-East, GOES-West, Meteosat 0, Meteosat Indian Ocean and Himawari into a single global mosaic. It spans 17 channels across visible, water vapor and infrared bands, and refreshes hourly even though native satellite scan schedules run between 10 and 30 minutes.
• What PARDIG is: PARDIG, short for Precipitation AI Reanalysis: Densely Inferred GPM, is a global rainfall map produced by a model trained on sparse radar observations from NASA’s GPM Core observatory. Its inputs are SatWeave imagery and atmospheric fields from the European Centre for Medium-Range Weather Forecasts.
• The specifications: DeepMind put PARDIG’s temporal resolution at 15 minutes, its spatial resolution at about 10 kilometers, and its average latency at roughly 5 hours. The release pairs a live feed with a historical archive running from 1999 to 2026, while SatWeave’s archive extends more than a decade.
• How it was tested: The company said PARDIG was evaluated against three continental radar networks, in the United States, Europe and Japan, and against more than 36,000 weather stations worldwide. That benchmarking set is the basis for the performance claims it attached to the release.
• The performance claim: DeepMind said PARDIG outperforms leading global references across all analyzed precipitation rates, including NASA’s IMERG-Final product, and improves detection skill for light-to-moderate hourly rainfall by more than 30% against that same benchmark. It also described the system as “particularly effective at capturing extreme precipitation.”
• Where to get it: Live and historical data for both datasets sit on Google Cloud Storage, with PARDIG’s historical catalogs also published on Google Earth Engine. The announcement points developers to WeatherNext documentation and to a PARDIG paper posted on arXiv under identifier 2611.13941.
• Unstated terms: DeepMind did not state file sizes for either dataset, and did not set out licensing terms for their use. No pricing or access tier was specified for the live feeds.
• Who built it: The work was credited to teams across Google DeepMind and Google Research, working with Google Cloud and Google Geo. DeepMind acknowledged data providers including ECMWF, NASA, JAXA, NOAA, EUMETSAT, Korea’s KMA, Japan’s JMA, the Netherlands’ KNMI and Synoptic Data.

Background:

WeatherNext 3 is DeepMind’s recently released AI forecasting model. Opening its training inputs shifts the release from a finished model to the raw material other developers can build on.

Between the lines:

The evaluation leans on dense radar coverage in the United States, Europe and Japan — regions where ground truth is strongest. For Gulf meteorological agencies, the value sits less in the benchmark scores than in the 1999–2026 archive: a continuous global rainfall record at 10-kilometer resolution is the kind of input local models normally lack. The absence of stated licensing terms leaves commercial use undefined for now.

What’s next

Watch for licensing terms and file-size details to appear alongside the Cloud Storage and Earth Engine catalogs, and for independent validation of the 30% detection-skill claim against IMERG-Final in regions outside the three radar networks tested.