Skip to content
NewsResearch

Google DeepMind WeatherNext 3 Ships Hourly 5-Kilometer Forecasts and Goes Live in Maps Platform, BigQuery, and Earth Engine

· by Pondero Newsdesk

The short version

Google DeepMind released WeatherNext 3 on September 3, 2026, generating hourly surface-variable forecasts at 5-kilometer resolution. Developers can query the model today through the Maps Platform Weather API, BigQuery, Google Earth Engine, and Cloud Storage.

Google DeepMind WeatherNext 3 Ships Hourly 5-Kilometer Forecasts and Goes Live in Maps Platform, BigQuery, and Earth Engine

Global AI weather forecasts at 5-kilometer resolution and one-hour update cycles are now queryable through Google's developer APIs, replacing the 25-kilometer, 6-hour-cadence WeatherNext 2 that previously powered Google Search, Maps, and the Gemini app.

What

Google DeepMind and Google Research published WeatherNext 3 on September 3, 2026, per the Google DeepMind blog. The model generates hourly surface-variable forecasts at 5-kilometer resolution, a roughly five-times improvement over WeatherNext 2's 25-kilometer grid and 6-hour cadence. Other surface variables run at 10-kilometer resolution; atmospheric variables such as wind speed remain at 25 kilometers.

WeatherNext 3 ingests live geostationary satellite mosaics updating every hour, which allows it to bypass the six-hour observation-assembly lag that traditional numerical weather prediction systems require before each model run. In independent live evaluations by Brightband, the model delivered precipitation forecasts up to 60% more accurate against the IMERG satellite benchmark and up to 30% more accurate against MRMS radar measurements, per the announcement.

Developers can access WeatherNext 3 through four channels starting September 3: the Google Maps Platform Weather API, BigQuery, Google Earth Engine, and Google Cloud Storage in Zarr format. Google also replaced WeatherNext 2 with WeatherNext 3 in Google Search, the Gemini app, and Google Maps on the same date.

Why it matters

The one-hour lag cycle is the sharpest change. Traditional numerical weather prediction systems (ECMWF, NOAA GFS) require roughly six hours to assemble observation files before each run. WeatherNext 3 cuts that floor to one hour by pulling live satellite data continuously. For teams building renewable-energy dispatch tools, logistics platforms that price weather risk, or precision agriculture applications, that lag reduction opens a tier of forecast accuracy that was not available through a publicly queryable API before this release.

The BigQuery and Earth Engine integrations have a direct practical implication for data teams: both services let analysts run queries against WeatherNext 3 output in place, without building custom pipelines to pull and normalize forecast files. The Zarr format in Cloud Storage gives Python-based workflows a direct route via xarray.

The 5-kilometer key-surface grid also changes what is possible at city and sub-regional scale. At 25 kilometers, a single grid cell covers an area roughly the size of a mid-size city. At 5 kilometers, district-level or corridor-level queries become meaningful. That matters for hyperlocal products in insurance, construction scheduling, and event management.

What to watch next

Google has not specified open-API pricing beyond standard Maps Platform and BigQuery rates, and has not indicated whether third-party weather platforms can access WeatherNext 3 output outside Google Cloud. A peer-reviewed technical paper with model architecture and training details has not been released as of publication. If Brightband or other independent evaluators publish a full benchmark comparison against ECMWF HRES and GraphCast, those results will serve as a more complete picture of where WeatherNext 3 sits in the operational forecast landscape.

Sources