Google DeepMind Launches WeatherNext 3 With Hourly 5-Kilometer Forecasts

Google DeepMind and Google Research announced the launch of WeatherNext 3 on September 3, 2026, a new-generation AI model for weather forecasting. The model predicts conditions on an hourly basis instead of the previous six-hour intervals, and delivers key surface variables, temperature and humidity, at resolutions of up to 5 kilometers.

What Changed

The biggest difference from its predecessor lies in how it was trained. WeatherNext 3 is the first global AI weather forecasting model to learn directly from raw satellite observations, rather than relying solely on data already processed by classical numerical weather prediction models. The model ingests hourly mosaics from geostationary satellites across 11 channels at 0.1-degree resolution, along with data from ground stations in the METAR and Mesonet networks and ship and buoy measurements from the ICOADS database going back to 2001.

It also draws on precipitation data from NASA's IMERG satellite system and Google's own PARDIG dataset. Combining these sources is meant to give the model a far denser picture of the atmosphere than earlier versions, which relied mainly on already-processed meteorological data.

Performance Against Competitors

According to Google DeepMind, WeatherNext 3 came out on top in the Operational WeatherBench benchmark, outperforming both rival AI models from Microsoft and Nvidia and traditional forecasts from national weather services, including the US National Weather Service and Europe's ECMWF. Over a one-week forecast horizon, the model shows an average 10 percent improvement over ECMWF's AIFS ENS v2 system for upper-atmosphere variables.

The biggest improvement shows up in precipitation forecasts, where the CRPS error metric dropped by as much as 60 percent against the IMERG satellite dataset and 30 percent against the MRMS radar product. For surface temperature, the improvement over WeatherNext 2 reaches around 30 percent. Google acknowledges the model still has weaknesses, including hexagon-shaped spatial artifacts stemming from the computational grid structure and discontinuities in forecasts at the boundaries of six-hour windows.

Applications for Energy and Agriculture

The new model predicts wind speed at 100 meters, the height of typical wind turbines, as well as cloud cover and solar radiation needed to estimate renewable energy output. Google says the more accurate, hourly forecasts should help grid operators and power market participants better plan for balancing output from wind and solar farms, where even small forecast errors translate into reserve capacity costs.

The second area Google points to is agriculture, particularly in developing regions, where more accurate precipitation and temperature forecasts can help with irrigation planning and protecting crops from extreme weather events.

Where the Model Will Be Available

WeatherNext 3 integrates immediately into Google's consumer products, Search, the Gemini app and Google Maps, as well as into the Maps Platform Weather API and Google Earth Engine for developers and researchers. Google says key variables from the model will, for the first time, directly power weather features across multiple products simultaneously, rather than being limited to select experimental apps.

This will be the first time some of the key variables are feeding and powering multiple Google products at once. - Samier Merchant, senior engineer at Google

The model is already available operationally, according to the technical paper published alongside the launch describing its architecture and benchmark results.

For users in Poland, this means in practice more accurate hourly forecasts in Google Maps and Search, though Google has not said when full 5-kilometer resolution will cover Central Europe to the same extent as the regions already tested. The development of AI-based weather models also matters for Poland's energy sector, where the growing share of wind and solar farms in the power mix requires more precise output forecasting for the grid operator.