Google DeepMind and Google Research have released WeatherNext 3, a global AI weather model that refreshes every hour and resolves key surface variables down to about 5 kilometers. According to Google, independent live evaluations by Brightband rank it as the company’s most accurate global weather model to date, and it is already rolling into Search, the Gemini app, Maps, and Google’s cloud data services.
The release targets two long-standing limits of AI forecasting: coarse local detail and a roughly six-hour wait for traditional numerical weather prediction analysis. WeatherNext 3 addresses both by ingesting a live global geostationary satellite mosaic directly and by training against raw weather station observations, not only smoothed reanalysis grids.
What WeatherNext 3 changes
Most AI weather models initialize from numerical weather prediction analysis, which carries a data lag of about six hours. WeatherNext 3 adds live satellite imagery as a direct model input, allowing a new global forecast 24 times a day. For fast-developing convection, storms, and precipitation bands, an hourly refresh grounded in current observations is a practical shift, not just a benchmark gain.
The second change is ground truth. Instead of learning only from reanalysis, WeatherNext 3 trains dedicated output heads directly on station measurements, including METAR airport stations, regional Mesonet networks, and ICOADS ship and buoy records. Its 5 km temperature and dew point fields are therefore calibrated to what instruments record on the ground, which matters most where terrain varies quickly, such as coastlines, valleys, and mountains.
The model itself is a Functional Generative Network mesh transformer, building on the probabilistic approach introduced with WeatherNext 2, with a larger latent size and deeper mesh. It produces 64-member ensemble forecasts, giving a spread of possible outcomes rather than a single deterministic answer.
How WeatherNext 3 compares with WeatherNext 2
Google describes WeatherNext 3 as roughly five times sharper than its predecessor. The headline 5 km figure applies to specific station-trained variables, while other fields remain coarser by design. One forward pass produces three tiers:
- 1. 0.05 degrees (about 5 km): station-trained 2 m temperature and dew point.
- 2. 0.1 degrees (about 10 km): gridded surface wind at 10 m and 100 m, pressure, sea surface temperature, cloud layers, solar radiation, and 1-hour precipitation.
- 3. 0.25 degrees (about 25 km): atmospheric fields across 13 pressure levels.
Cadence has changed just as much as resolution. WeatherNext 2 issued forecasts in 6-hour increments. WeatherNext 3 runs every hour, with the 00, 06, 12, and 18 UTC synoptic cycles extending to 15 days and 64 members, while interim hourly runs cover 48 hours.
Readers following Google’s broader time-series work may also be interested in our coverage of Google’s TimesFM-3 multivariate forecasting model, which tackles a different but related prediction problem.
Which AI is the best for predicting weather?
There is no single winner for every variable, region, and lead time. Google points to Brightband’s Operational WeatherBench, where WeatherNext 3 led models from Microsoft, Nvidia, and ECMWF as well as traditional forecasts from the US National Weather Service on tested temperature, wind, and humidity metrics. Independent verification across seasons and extremes will still matter, and users should treat vendor-reported skill scores as a starting point rather than a final verdict.
What is Google DeepMind WeatherNext 2?
WeatherNext 2 was Google’s previous operational AI forecaster, producing 0.25 degree global fields in 6-hour steps. It powered earlier weather experiences in Google products and cloud datasets. WeatherNext 3 replaces it in Search, Gemini, and Maps, while on-demand custom inference still runs on WeatherNext 2 for now.
Why precipitation and clean energy forecasts matter
Precipitation is where global models often fail, blurring rain into diffuse fields that miss storm edges. WeatherNext 3 trains against three precipitation sources: ECMWF reanalysis, NASA’s IMERG satellite retrievals, and Google’s own satellite-radar precipitation reanalysis. Google reports continuous ranked probability score improvements of up to 60% against IMERG, 30% against MRMS radar, and 10% against rain gauges at early lead times, with up to 50% better precipitation forecasts in consumer products a day or more ahead.
For clean energy, the model outputs 100 m wind speed at approximate turbine hub height, full low, medium, and high cloud distributions, and both solar irradiance components. That combination helps grid operators anticipate wind and solar output against demand, a direct link between forecast skill and renewable integration.
The technical documentation is published in Google’s WeatherNext developer guides, with the full methods in the WeatherNext 3 research paper.
Where you will see it and how to access the data
Google says WeatherNext 3 has begun powering weather experiences in Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine. For developers and researchers, hourly forecast grids are available through BigQuery, Earth Engine, and Cloud Storage in Zarr format after an allowlist request, with no model setup required. Google also points researchers to Weather Lab for experimentation.
Access is not the same as open source. Forecast data is available by request, but WeatherNext 3 weights are not open, and Google has not detailed pricing beyond standard Maps Platform and BigQuery rates. As TechCrunch reported, Google engineer Samier Merchant called this the first time core model variables will directly power Google products, which makes independent evaluation more important as adoption spreads.
What remains limited
Three caveats deserve attention. First, the 5 km resolution covers temperature and dew point, not every variable, so readers should not assume all fields are equally sharp. Second, performance claims so far come largely from Google’s evaluations and the Brightband leaderboard it cites, and separate comparisons across regions, seasons, and extreme events have yet to accumulate. Third, experimental AI forecasts are not official warnings, and fast-moving hazards still require national meteorological services. For background on Google’s official release, see the WeatherNext 3 announcement on the Google blog.
