WeatherNext 3 goes hourly: what DeepMind's latest AI weather model means for builders and operators

Google DeepMind released WeatherNext 3, a global weather AI model that refreshes forecasts every hour using real-time satellite imagery rather than waiting for government datasets that update every six hours. For builders, this means access to fresher, higher-resolution weather data through Google Search, Maps, Gemini, and cloud services like BigQuery and Earth Engine, with specific value for renewable energy forecasting and cyclone tracking.

Hourly forecasts from live satellite data

Most AI weather models depend on the European Centre for Medium-Range Weather Forecasts (ECMWF) dataset, which takes about five hours to produce and is refreshed only every six hours. WeatherNext 3 layers real-time geostationary satellite imagery on top of that ECMWF output, generating predictions every hour and reducing the data lag to roughly three to four hours. As DeepMind senior research scientist Ilan Price told Bloomberg, "It gets much more accurate by not waiting for the next analysis date and using the most recent information."

What WeatherNext 3 improves

The model boosts resolution substantially. Surface variables like temperature and moisture are visualized on a 5-kilometer grid, compared to the 25-kilometer grid used by WeatherNext 2. Wind speed forecasts still run at 25-kilometer resolution, but the model can target forecasts to specific weather stations rather than averaged grid cells. On precipitation, DeepMind reports up to 60% better probabilistic accuracy than the previous version, trained on NASA's satellite-based precipitation dataset and a Google-produced precipitation reanalysis. The model also has 2.4 times more parameters than WeatherNext 2.

Where you can access the data

WeatherNext 3 now powers weather results across Google Search, Gemini, Google Maps, and the Google Maps Platform Weather API. Developers and researchers can access forecast data through BigQuery, Google Earth Engine, and Google Cloud Storage. That means you can pull hourly weather predictions directly into your data pipelines or geospatial analyses without building custom ingestion from public datasets. The integration into Google's consumer products also means users get more accurate, more local weather information without needing to switch apps.

The renewable energy angle

The hourly cadence is especially useful for wind and solar energy operators. WeatherNext 3 forecasts wind speeds at 100 meters (turbine height), cloud cover, and solar radiation. Grid operators and renewable energy developers can use that data to estimate power output from wind and solar assets with fresher inputs. The model also natively predicts cyclone tracks alongside its gridded atmospheric outputs, which could improve disaster preparedness workflows.

What remains to be verified

DeepMind says WeatherNext 3 performed well in ongoing head-to-head testing on Brightband's Operational WeatherBench platform, which evaluates models across temperature, wind speed, and humidity. However, the model's forecasts have not been independently verified beyond those evaluations. Performance claims around precipitation accuracy and cyclone prediction are vendor-reported, and users should treat them as directional until third-party studies confirm the results. The model is rolling out now; if you rely on weather data for critical operations, test its outputs against your own ground truth before depending on it entirely.