Hourly, multi-resolution grids replace WeatherNext 2's 25-kilometer, six-hour picture. Google DeepMind and Google Research posted WeatherNext 3 on 3 September 2026.
The news
On 3 September 2026, Google DeepMind and Google Research introduced WeatherNext 3. The post calls it the most advanced and accurate global weather model to date, citing independent live evaluations by Brightband. The model learns from real-time observations, including raw geostationary satellite data, and produces a new high-resolution forecast every hour.
Resolution stated: key surface variables such as temperature and moisture at 5-kilometer resolution; other surface variables at 10 kilometers; atmospheric variables such as wind speed at 25 kilometers. The post says that is roughly five times sharper than WeatherNext 2, which produced forecasts on a 25-kilometer grid in 6-hour increments. Architecture sketch: live 1-hour geostationary satellite mosaics plus historical analysis into a Functional Generative Network mesh transformer; outputs include dense gridded fields, discrete cyclone tracks, and station-level sparse coordinates.
Training shift: most AI weather models, including WeatherNext 2, train on numerical weather prediction fields with about a six-hour lag. WeatherNext 3 ingests live global geostationary satellite mosaics and trains directly on sparse weather-station observations for local detail. Precipitation training sources named: NASA IMERG and Google's own global precipitation reanalysis based on satellite radar.
Company evaluation claims: medium-range precipitation CRPS improvement of up to 60% against IMERG, 30% for MRMS, and 10% against rain gauges for early lead times. Product claim: when planning a day or more ahead, people will see up to 50% more accurate precipitation forecasts, with the largest gains where forecasts were historically less reliable. Those percentages are Google's.
Clean-energy variables named: 100-meter wind speeds for turbine height, high-resolution cloud cover, and sun radiation for solar. Distribution: query data in BigQuery and Earth Engine, or bulk-download from Google Cloud Storage; the WeatherNext developer docs are linked. Product surfaces starting the same day: Google Search, Gemini app, Google Maps, Google Maps Platform Weather API, and Google Earth Engine.
Paper link on the post: arXiv 2609.03582. Brightband live leaderboards and Weather Lab are linked. Disclaimer on the post: for official forecasts, severe warnings, and public safety advisories, refer to your local meteorological agency or national weather service.
Who is bound
Google DeepMind and Google Research shipped the model and the product integrations. Google Search, Gemini, Maps, Maps Platform, Earth Engine, BigQuery, and Cloud Storage customers see the new fields where rolled out. Brightband is the named independent eval host, not a Google unit on this page. National weather services remain the safety authority per Google's own disclaimer.
What's new
An hourly, observation-grounded global model at 5-kilometer surface detail, with renewables variables and same-day wiring into consumer Google weather surfaces plus Cloud data paths. It is not Gemini 3.8 Flash. It is not Lyria. WeatherNext 2 open-sourcing and cyclone papers are prior line items, not this launch.
What it does not settle
Date: 3 September 2026. Brightband methodology beyond the linked leaderboard is UNKNOWN here. List price for bulk Cloud weather data is UNKNOWN on this page. Exact country-by-country Maps and Search cutover timing is UNKNOWN beyond starting today. Operational replacement of a national weather service is explicitly disclaimed. Hardware to self-host weights is UNKNOWN; the post pushes managed data access, not a public checkpoint dump on this page.
What to do now
Maps Platform and Cloud weather API buyers should check the WeatherNext developer docs for field names and refresh cadence before rewriting pipelines still on six-hour grids. Energy desks that need 100-meter wind should verify the variable in BigQuery or Earth Engine, not assume a turbine SCADA feed. Emergency managers should keep national weather service warnings as the authority Google already points to. Ignore most accurate as a purchase order without the Brightband slice you care about.