Google DeepMind and Google Research have introduced WeatherNext 3, which the company describes as its most advanced global weather forecasting model to date. Independent live evaluations conducted by Brightband indicate the system delivers a significant leap in predictive accuracy and spatial resolution relative to both conventional numerical methods and its predecessor, WeatherNext 2.

The new model breaks from established practice by learning directly from real-time observations instead of relying solely on historical numerical weather prediction data, which typically carries a six-hour lag. WeatherNext 3 ingests live global geostationary satellite mosaics every hour, allowing it to generate hourly forecasts at resolutions as fine as five kilometers for surface variables such as temperature and moisture. The shift toward AI-based forecasting matters well beyond Google's own products: machine-learning forecast models generally run at a fraction of the computational cost of traditional supercomputer-based numerical weather prediction, a factor that shapes how quickly agencies such as the NOAA, ECMWF, and national meteorological services adopt comparable techniques.

That represents a fivefold improvement in spatial precision over WeatherNext 2, which operated on a 25-kilometer grid with six-hour increments; atmospheric variables including wind speed are resolved at 25 kilometers. At the core of the system is a Functional Generative Network mesh transformer that processes both satellite imagery and traditional historical analysis to natively produce dense gridded fields, discrete cyclone tracks, and station-level predictions.

Introducing WeatherNext 3, our most advanced and accurate global weather AI model to date, from @GoogleDeepMind and @GoogleResearch. This new forecasting model learns directly from real-time observations, and uses raw satellite data to produce a forecast every hour in high… pic.twitter.com/WlljsjiSfy — News from Google (@NewsFromGoogle) September 3, 2026

Energy Forecasting and Global Availability

Beyond technical refinement, WeatherNext 3 brings specialized capabilities for the renewable energy sector and historically underserved regions. The model forecasts 100-meter wind speeds together with high-resolution cloud cover and solar radiation estimates, giving grid operators and developers precise tools for predicting clean energy output and balancing supply against consumer demand. That capability addresses a longstanding operational challenge for grids with growing shares of wind and solar, where output variability requires accurate short-term forecasts to schedule dispatch and maintain reliability.

Because it trains directly on sparse weather station observations rather than smoothed atmospheric simulations, the system accounts for local topography and extreme microclimatic variations that are especially relevant to coastal, valley, and mountain communities. This methodology is particularly consequential for Latin America, Africa, and Asia-Pacific, where high-resolution forecasting has historically been limited by the prohibitive supercomputing costs of traditional regional models.

Precipitation accuracy has improved markedly through training on NASA's IMERG satellite data and Google's proprietary global precipitation reanalysis. Independent evaluations show Continuous Ranked Probability Scores up to 60% better against IMERG benchmarks, along with 30% improvements against MRMS and 10% against rain gauge measurements at early lead times.

Google is deploying WeatherNext 3 across its consumer and enterprise ecosystems with immediate effect. The model now powers weather experiences in Search, Gemini, Maps, the Google Maps Platform Weather API, and Earth Engine. Users planning travel or outdoor activities may encounter precipitation forecasts up to 50% more accurate at longer-range horizons, with the most pronounced gains in regions that previously lacked reliable forecasting infrastructure.

For researchers and businesses, hourly global prediction datasets can be queried through BigQuery and Earth Engine or downloaded in bulk via Google Cloud Storage without the need for dedicated model deployment. The release continues Google's line of WeatherNext models and its GraphCast work, part of a broader industry trend in which AI weather models from major technology firms and meteorological agencies are increasingly used alongside, rather than instead of, conventional physics-based forecasting systems.