Google DeepMind's WeatherNext 3 Delivers Hourly Forecasts at 5-Kilometer Resolution

Google DeepMind and Google Research have released WeatherNext 3, a global weather model that generates hourly forecasts at 5-kilometer resolution — roughly five times sharper than its predecessor — and which the company calls its most accurate global weather model to date, according to independent live evaluations by Brightband.

The release, announced today, marks a shift in how AI weather models are built. Most AI forecasters, including Google's own WeatherNext 2, train on outputs from numerical weather prediction (NWP) models — supercomputer-driven physics simulations that carry a six-hour data lag. That lag can bias fast-changing variables like rain and surface temperature. WeatherNext 3 instead learns directly from real-time observations, ingesting a mosaic of live global geostationary satellite data to produce a new forecast every hour, each grounded in the most recent satellite imagery available.

The stakes extend well beyond whether to grab an umbrella. Wind, rain, heatwaves and drought cascade across agriculture, global supply chains, clean energy production and national economies, and in recent years AI models have outpaced traditional methods in speed and accuracy. Yet predicting highly local, rapidly changing weather has remained difficult. Previous models lacked sufficient spatial resolution and struggled to incorporate real-time satellite data.

Sharper forecasts, faster updates

WeatherNext 3 visualizes key surface variables such as temperature and moisture at 5-kilometer resolution, other surface variables at 10 kilometers, and atmospheric variables like wind speed at 25 kilometers, maintaining physical consistency from broad global wind patterns down to local topography. WeatherNext 2, by comparison, produced forecasts on a 25-kilometer grid in 6-hour increments.

The update cycle matters because critical weather develops fast. When storms, fronts or precipitation systems materialize suddenly, Google says the rapid refresh and higher resolution provide earlier, more detailed insights needed to drive an effective response.

The model also trains directly on sparse weather station observation data rather than on coarse atmospheric representations. Temperature and humidity can fluctuate dramatically over just a few kilometers — particularly near coastlines, valleys and mountain ranges — and traditional models miss those extremes. Google frames this as especially significant for Latin America, Africa and Asia-Pacific, regions historically underserved by high-resolution forecasting because of the supercomputing costs of traditional regional models.

Architecturally, the system ingests live one-hour geostationary satellite mosaics alongside traditional historical analysis, feeding a single Functional Generative Network (FGN) mesh transformer that outputs dense gridded fields, discrete cyclone tracks, and station-level sparse coordinate predictions natively.

Breakthroughs in precipitation accuracy

Precipitation is a notorious weak point for global models. Rain and snow systems are driven by fast-moving cloud processes at tiny scales that physics-based simulations struggle to capture, and AI forecasts often return blurry estimates or miss severe storm boundaries entirely.

To counter this, Google trained WeatherNext 3 on two high-quality precipitation sources: NASA's satellite-based Integrated Multi-satellite Retrievals for GPM (IMERG) and its own global precipitation reanalysis based on satellite radar. In medium-range global forecasts, evaluations against baselines show a Continuous Ranked Probability Score (CRPS) improvement of up to 60% against IMERG, 30% for MRMS, and 10% against rain gauge measurements at early lead times. Side-by-side comparisons show WeatherNext 3 at 11-kilometer resolution closely mirroring satellite ground truth and capturing sharp convective bands that WeatherNext 2 smeared into diffuse, pixelated footprints.

Built for renewable energy

WeatherNext 3 also introduces predictions engineered specifically for renewable energy production. It forecasts 100-meter wind speeds — roughly turbine height — for wind-energy output, alongside high-resolution cloud cover and solar radiation levels so solar farms can estimate ground-level light. Google positions this data as crucial for grid operators and renewables developers to predict clean energy generation and match it with consumer demand.

Deployment starts today

WeatherNext 3 begins powering weather experiences across Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API and Google Earth Engine starting today. Google says people planning a day or more ahead will see up to 50% more accurate precipitation forecasts, with the largest gains in regions where forecasts have historically been less reliable.

For developers and researchers, Google is making hourly-updated global forecast data available through BigQuery and Earth Engine, with bulk downloads from Google Cloud Storage and no model setup required. The company has also published a paper, an API for building on the model, a Weather Lab visualization tool showing WeatherNext 3 in real time, and Brightband's independent live leaderboards tracking where the model ranks.

Google cautions that official severe weather warnings and public safety advisories still come from local meteorological agencies and national weather services — and that "the atmosphere will always retain a degree of unpredictability." But by training on real-world observations and bypassing traditional modeling constraints, the company argues WeatherNext 3 brings forecasting closer to matching what actually happens on the ground, with implications for emergency responders, air traffic controllers, farmers and grid operators alike.