AI NewsSep 4, 2026Words 1532Read time≈ 4 min

Google Launches WeatherNext 3 With Hourly Satellite-Informed Global Forecasts

WeatherNext 3 adds hourly initialization, outputs as fine as 5 km, 64-member ensembles, and new precipitation and clean-energy fields.

AIGoogle

Contents · 11

1. What Changed From WeatherNext 2
2. How Live Observations Enter the Model
3. Reading Google’s Accuracy Claims Correctly
4. Access, Energy Applications, and Operational Limits
Frequently Asked Questions
Does WeatherNext 3 update every hour?
Are all WeatherNext 3 forecasts produced at 5-kilometer resolution?
Is WeatherNext 3 open source?
Where can developers access the forecasts?
Can WeatherNext 3 replace official weather warnings?
Sources

Google DeepMind and Google Research introduced WeatherNext 3 on September 3, 2026, adding live geostationary satellite imagery, hourly forecast initialization, and surface outputs as fine as approximately 5 kilometers to Google’s global AI weather system.

The model is beginning to power weather experiences in Google Search, the Gemini app, Google Maps, Google Maps Platform’s Weather API, and Google Earth Engine. Researchers and businesses can also request access to operational forecast data through BigQuery, Earth Engine, and Google Cloud Storage.

WeatherNext 3 is a probabilistic ensemble model rather than a single deterministic forecast. It produces 64 possible atmospheric evolutions, allowing downstream systems to estimate uncertainty and the probability of events such as rainfall exceeding a specified threshold.

Its resolution gains are substantial but variable-specific. Weather-station-calibrated temperature and dew-point predictions reach 0.05 degrees, or roughly 5 kilometers. Gridded surface variables—including precipitation, wind, pressure, cloud cover, and solar radiation—use a 0.1-degree grid of about 10 to 11 kilometers. Upper-air variables remain on a 0.25-degree grid of approximately 25 kilometers.

1. What Changed From WeatherNext 2

WeatherNext 2 produced global forecasts on a 0.25-degree grid and was initialized four times daily, at six-hour intervals. WeatherNext 3 can be initialized every hour and generates outputs with native one-hour timesteps.

The new schedule has an important qualification. Forecasts initialized at 00:00, 06:00, 12:00, and 18:00 UTC extend through the full 360-hour, or 15-day, horizon. The 20 interim hourly runs cover surface and weather-station variables for 48 hours. Upper-atmosphere fields are available only for the four main six-hourly cycles.

Google’s statement that the new model is “up to five times sharper” refers to the change from WeatherNext 2’s 25-kilometer grid to WeatherNext 3’s 5-kilometer station-calibrated temperature and dew-point products. The 10-kilometer gridded surface products offer an improvement closer to 2.5 times, while the 25-kilometer pressure-level fields do not receive the same spatial-resolution increase.

This distinction matters for users evaluating local applications. A five-kilometer temperature product can represent coastlines, valleys, and terrain more precisely than a 25-kilometer grid, but it does not mean every atmospheric variable is forecast at five-kilometer resolution.

WeatherNext 3 also changes precipitation accumulation intervals. WeatherNext 2 reported six-hour accumulated precipitation, while the new model exposes native, IMERG-calibrated, and experimental satellite-radar precipitation fields in one-hour accumulations. Existing data pipelines must sum hourly values when calculating daily or multi-day totals.

2. How Live Observations Enter the Model

The central technical change is the direct ingestion of hourly mosaics assembled from geostationary weather satellites. These satellites continuously observe broad regions of Earth, giving the model a recent picture of cloud systems and other rapidly changing atmospheric patterns.

WeatherNext 3 does not eliminate numerical weather prediction data. Google’s developer documentation lists its initialization inputs as live satellite mosaics plus analysis from the European Centre for Medium-Range Weather Forecasts’ high-resolution system. Its training sources include ERA5 and operational ECMWF analysis, NASA’s Integrated Multi-satellite Retrievals for GPM, weather-station observations, and historical geostationary satellite mosaics.

The result is therefore a hybrid system: satellite data improves the recency of its view, while traditional analysis still provides part of the atmospheric state. This is more precise than saying the model bypasses physics-based forecasting altogether.

Those inputs feed a Functional Generative Network mesh transformer. A single model produces dense gridded fields, sparse station-targeted predictions, and discrete cyclone tracks while maintaining consistency between global atmospheric patterns and local surface outputs. TechCrunch reported that the model has 2.4 times as many parameters as WeatherNext 2.

Dedicated observational output heads address a long-standing limitation of models trained only against reanalysis grids. WeatherNext 3 learns temperature and dew point directly from weather-station measurements, letting it optimize predictions toward quantities recorded at actual stations rather than only approximating values in a smoothed global grid.

The model applies the same principle to precipitation through multiple targets. It trains against ECMWF reanalysis, NASA IMERG satellite estimates, and a Google-developed satellite-radar precipitation reanalysis. This produces separate precipitation fields rather than treating one dataset as an unquestioned ground truth.

3. Reading Google’s Accuracy Claims Correctly

Google reports improvements in the Continuous Ranked Probability Score, or CRPS, of up to 60% when WeatherNext 3 precipitation forecasts are evaluated against IMERG, 30% against the US Multi-Radar/Multi-Sensor system, and 10% against rain-gauge measurements at early lead times.

CRPS evaluates an entire probability distribution rather than only one predicted value; lower scores indicate that forecast probabilities are both sharper and closer to observations. The percentages therefore describe relative improvements in a probabilistic evaluation metric. They do not mean that WeatherNext 3 predicts rain correctly 60% more often in every location.

The spread between the three results is also informative. Satellite estimates, radar products, and ground gauges observe precipitation differently and have different geographic coverage and biases. A model trained partly against IMERG should not be expected to show identical gains when judged against independent rain gauges.

Google separately says that people checking forecasts at least one day ahead in its consumer products may see precipitation predictions become up to 50% more accurate, with the largest gains in regions where forecasts have historically been less reliable. The launch announcement does not provide enough methodological detail to equate that product-level figure with the reported CRPS results.

Independent live evaluation offers another reference point. TechCrunch reported that WeatherNext 3 led Brightband’s Operational WeatherBench at launch across assessed variables including temperature, wind speed, and humidity. Because the leaderboard evaluates operational forecasts continuously, its ordering can change as new forecasts and competing systems are added.

There is also a boundary around Google’s claim about using raw observations. WeatherNext 3 directly incorporates satellite observations at global scale, but TechCrunch noted that WindBorne says its WeatherMesh 6 system has ingested raw balloon and other observations since 2025. Google’s narrower differentiator is the combination of direct satellite ingestion, global coverage, and higher-resolution output—not an uncontested claim to being the first weather model to use any raw observations.

4. Access, Energy Applications, and Operational Limits

WeatherNext 3 exposes variables intended for renewable-energy forecasting as first-class outputs. These include wind speed at 100 meters, approximately turbine-hub height, multiple cloud-cover layers, surface downward solar radiation, and direct-beam solar radiation. Wind and solar operators can use those fields to estimate generation and balance expected supply against demand.

The publicly documented Earth Engine products contain precomputed ensemble statistics. The 0.05-degree collection provides the mean and the 10th, 25th, 50th, 75th, and 90th percentiles for station-targeted temperature and dew point. The 0.1-degree collection provides the same six statistics for 19 gridded surface variables. The full 64-member ensemble and upper-air fields are distributed through Google Cloud Storage.

These datasets require an access request. WeatherNext 3 itself is available as an operational forecast service but is not open source; Google instead recommends WeatherNext 2 to researchers who need downloadable code and pretrained weights for self-hosted inference.

Operational availability also has latency. Google’s dissemination schedule targets delivery of the four full 15-day cycles to Cloud Storage about seven hours and 45 minutes after initialization, followed approximately 25 minutes later by BigQuery and Earth Engine. Interim hourly runs target Cloud Storage delivery roughly seven hours and 10 minutes after initialization and BigQuery or Earth Engine about 15 minutes after that. Upstream satellite or analysis delays can extend those windows.

Google labels WeatherNext 3 an experimental forecasting system. Its outputs are supplied for informational and research use, not as official watches or warnings, and should not be the sole source for decisions involving life or property. Historical experimental data is licensed under CC BY 4.0, while real-time and future forecast data is governed by separate Google DeepMind terms.

Frequently Asked Questions

Does WeatherNext 3 update every hour?

Yes. The model is initialized 24 times per day and produces one-hour forecast timesteps. Four daily runs extend to 15 days; the interim hourly runs cover 48 hours.

Are all WeatherNext 3 forecasts produced at 5-kilometer resolution?

No. The approximately 5-kilometer output applies to station-calibrated temperature and dew point. Most gridded surface variables use roughly 10-kilometer resolution, while pressure-level atmospheric fields use approximately 25 kilometers.

Is WeatherNext 3 open source?

No. Google provides it as an operational forecast-data service. WeatherNext 2 remains the recommended open-source option for users who need code and pretrained weights.

Where can developers access the forecasts?

Google distributes WeatherNext 3 data through BigQuery, Earth Engine, and Google Cloud Storage after an access request. Consumer forecasts are also beginning to appear in Search, Gemini, Maps, and the Google Maps Platform Weather API.

Can WeatherNext 3 replace official weather warnings?

No. Google explicitly classifies the system as experimental and directs users to national meteorological agencies and local emergency authorities for official forecasts, watches, and warnings.