Introducing WeatherNext 3, our most advanced and accurate global weather AI model

Google DeepMind and Google Research's own blog introduces WeatherNext 3 as their most advanced global weather model, citing independent live evaluations by Brightband. The headline shift is training on real-time observations rather than only numerical weather prediction output. The model ingests live one-hour geostationary satellite mosaics plus historical analysis through a Functional Generative Network mesh transformer, and emits gridded fields, cyclone tracks, and station-level coordinates.

Resolution: key surface variables such as temperature and moisture at 5 kilometers, other surface variables at 10 kilometers, atmospheric variables such as wind at 25 kilometers — described as roughly five times sharper than WeatherNext 2's 25-kilometer grid and six-hour steps. Forecasts refresh hourly. Precipitation training uses NASA IMERG and Google's own satellite-radar reanalysis; Google reports CRPS improvements of up to 60 percent against IMERG, 30 percent for MRMS, and 10 percent against rain gauges at early lead times. Users planning a day or more ahead are told to expect up to 50 percent more accurate precipitation forecasts, with the largest gains where forecasts have been weakest. The model also predicts 100-meter wind, cloud cover, and solar radiation for energy. Data is going into Search, Gemini, Maps, Maps Platform Weather API, Earth Engine, BigQuery, and Cloud Storage. Google still tells readers to use national weather services for warnings.

This is a vendor blog. Brightband is named as an independent live leaderboard, which is the one external check on the page. The CRPS stack (60 / 30 / 10) is Google's evaluation against its chosen baselines. Treat 'most advanced and accurate to date' as a press-release superlative until those leaderboards are read without Google's caption.

The architectural claim that matters is skipping the six-hour NWP lag by training on live satellite mosaics and sparse station data. That is how you get hourly, 5-kilometer temperature fields in places that never paid for a regional supercomputer. Latin America, Africa, and Asia-Pacific are named for a reason: that is both a development story and a Maps/Search distribution story.

Signal Desk's read: WeatherNext 3 is Google putting a forecast model inside the consumer surface billions already open, then selling the same fields as Cloud data. The 50 percent day-ahead precipitation number will be the one that leaks into every rewrite. It is a product metric, not a warning-system replacement — Google's own disclaimer says so. The energy variables (100-meter wind, irradiance) are not altruism. Google is a power buyer. Better renewable forecasts are also a data-center procurement input.

Hourly satellite in, Search out, is a distribution coup traditional weather agencies cannot match. Accuracy is now a product feature Google can ship on a blog cadence.

AI weather models trained on ECMWF-style reanalyses have been catching physics-based global models for several years. WeatherNext 2 still ran on a 25-kilometer, six-hour cadence like those systems. Version 3 is Google's attempt to ingest live observations directly and put the result in consumer apps the same day.