Google DeepMind launches WeatherNext 3, its sharpest global weather AI model yet

Google DeepMind and Google Research introduced WeatherNext 3 on September 3, 2026, a global weather forecasting model the companies say is now the most accurate available according to independent evaluators, and rolled it out immediately across Search, Maps, Gemini, and Google Cloud.

What's new

WeatherNext 3 is, in Google's words, "the most advanced and accurate global weather model to date, according to independent live evaluations by Brightband." The model produces hourly forecasts at multiple spatial resolutions rather than the coarser, less frequent output of its predecessor:

• Surface variables like temperature and moisture: 5-kilometer resolution
• Other surface variables: 10-kilometer resolution
• Atmospheric variables such as wind speed: 25-kilometer resolution

Google says: "Overall, this provides a global weather picture roughly five times sharper than our previous model, WeatherNext 2, which produced forecasts on a 25-kilometer grid in 6-hour increments."

The jump in resolution and update frequency comes from a change in inputs: instead of relying solely on traditional numerical weather prediction, WeatherNext 3 ingests "a mosaic of live, global geostationary satellite data," which is what enables hourly updates at up to 5-kilometer resolution.

Precipitation forecasting shows the clearest gains. Google reports "a Continuous Ranked Probability Score (CRPS) improvement of up to 60% against IMERG, 30% for MRMS, and 10% against rain gauge measurements for early lead times," and says medium-range forecasts are "up to 50% more accurate" for precipitation specifically.

The model also adds two new outputs aimed at the energy sector: predicted wind speeds at 100 meters, the typical hub height for wind turbines, and high-resolution cloud cover and solar radiation data for solar planning.

Context

WeatherNext has been Google DeepMind's testbed for applying AI to weather prediction since its first version, with WeatherNext 2 previously highlighted for improved cyclone-track forecasting. Each iteration has pushed toward higher resolution and lower latency than physics-based numerical weather models, which are computationally expensive and typically update on a fixed, coarser cadence. WeatherNext 3's move to satellite-driven hourly updates marks a structural shift away from pure simulation toward a hybrid, observation-heavy approach.

Google is making the model available immediately and broadly: WeatherNext 3 is live "across Google Search, Gemini app, Google Maps, Google Maps Platform Weather API, and Google Earth Engine starting today," and developers can pull the raw data via BigQuery, Earth Engine, or bulk downloads from Google Cloud Storage.

Why it matters

Weather forecasts feed directly into consumer products people check daily — Search snippets, Maps trip planning, Gemini queries — as well as into commercial planning for agriculture, logistics, and energy. A model that updates hourly at 5-kilometer resolution instead of every six hours at 25 kilometers changes what's practical to build on top of it: same-day localized planning, more responsive renewable-energy dispatch, and tighter short-range precipitation warnings. By shipping WeatherNext 3 simultaneously across its consumer surfaces and its developer/cloud tooling, Google is positioning weather AI as infrastructure it controls end to end, from the raw satellite-fed model down to the Maps widget a user glances at before leaving the house.