WeatherNext 3 updates its global forecasts every hour

Google is rolling out WeatherNext 3, a global weather model that uses recent satellite observations, updates every hour and reaches a resolution of about three miles for selected variables.

A new global forecast can now begin every hour, instead of waiting for one of the four main daily weather cycles. With WeatherNext 3, Google is bringing its model closer to the latest observations while refining selected surface data to a resolution of about three miles.

Introduced on September 3, 2026, WeatherNext 3 succeeds WeatherNext 2, which mainly produced forecasts at a 15-mile resolution and in six-hour increments. The new version can generate hourly estimates covering the next 15 days, with complete updates every six hours and interim hourly runs for the first 48 hours.

The improvement does not mean the entire atmosphere is uniformly converted into a three-mile grid. Google uses several levels of detail depending on the variable. Station-calibrated temperature and dew point are available on a 0.05-degree grid, equivalent to roughly three miles. Core surface variables are calculated at 0.1 degrees, or about six miles. Atmospheric data distributed across 13 pressure levels remains at 0.25 degrees, close to 15 miles.

This distinction matters because the highest resolution promoted by Google does not apply to every output. Nor does it turn a global forecast into a street-level observation. A three-mile cell can still contain different terrain, urban areas, bodies of water and several microclimates that the model does not represent separately.

WeatherNext 3 produces probabilistic forecasts. For each lead time, it generates an ensemble of 64 possible scenarios rather than a single future treated as certain. Their spread provides an estimate of uncertainty: when the members converge, the situation is relatively stable; when they diverge, the forecast calls for greater caution.

This approach is particularly useful for rainfall, storms and tropical cyclone tracks, where small differences in initial conditions can significantly alter what happens next. Developer-facing interfaces can provide an average, a median or several quantiles. The complete data hosted in Google Cloud Storage also allows users to examine each of the 64 ensemble members separately.

The model covers temperature, humidity, pressure, wind at 10 and 100 meters, precipitation, cloud cover, solar radiation, sea surface temperature and several atmospheric variables. Dedicated outputs are also available for weather stations and tropical cyclone tracking.

The main change lies in the data received at the beginning of each forecast. WeatherNext 3 uses two atmospheric states separated by six hours, together with the 12 most recent hourly satellite mosaics. These mosaics combine 11 channels from geostationary satellites to provide an almost continuous view of clouds, water vapor and other atmospheric features.

The newest satellite data arrives with a delay of just under one hour, while the conventional atmospheric analyses used by the system can be about five hours behind real time. Interim runs therefore use recent satellite observations to bring the forecast closer to current conditions, even when a new complete atmospheric analysis is not yet available.

This structure explains the hourly updates announced by Google. The model is not retrained from scratch every hour. Instead, a new run begins using the latest available inputs. Interim forecasts cover 48 hours, while those initialized every six hours extend to 360 hours, or 15 days.

WeatherNext 3 does not rely exclusively on the atmospheric archives commonly used to train weather models. It continues to use ERA5 and the European Centre for Medium-Range Weather Forecasts’ operational HRES analyses, while adding several sources of more direct observations.

The training data includes readings from about 5,000 METAR airport stations, alongside nearly 15,000 Mesonet stations available in 2024. Measurements collected by ships and buoys through the ICOADS database supplement these sources. Google held back 5% of the land stations to evaluate the model’s ability to estimate conditions at locations it had not encountered during training.

For precipitation, Google uses NASA’s IMERG satellite observations and PARDIG, an internal reanalysis that combines satellite and radar data. The model has a dedicated output trained to match these observations, in addition to rainfall estimates derived from conventional weather analyses.

The use of several references addresses a particular challenge: determining how much rain actually fell remains difficult at a global scale. Satellites, radar systems, rain gauges and reanalyses do not measure exactly the same thing and carry different biases. A model can therefore receive a different score depending on the source chosen for evaluation.

In its technical paper, Google reports reductions in probabilistic error of up to 60% compared with WeatherNext 2 and the European ensemble when forecasts are evaluated against IMERG at short lead times. The improvement reaches up to 30% against the US MRMS radar system and about 10% against rain gauges during the first few hours.

The largest gains reportedly appear around the distinction between rain and no rain. Google also says users could receive precipitation forecasts that are up to 50% more accurate at lead times of one day or longer. That percentage still depends on the region, forecast horizon, observation reference and statistical measure being used. It cannot be applied uniformly to every rainfall event.

Hourly initialization itself reportedly provides the equivalent of two to three additional hours of useful lead time for short-range precipitation forecasts. The model can then use more recent satellite imagery than was available for the previous run. The benefit should be greatest during the first few hours and decline as the forecast extends further into the future and atmospheric uncertainty becomes more influential.

The station-calibrated output also improves forecasts near the surface. On stations withheld from training, Google measured reductions in probabilistic error of up to 30% for two-meter temperature compared