Google DeepMind Launches WeatherNext 3 With Hourly 5-Kilometer Forecasts

Google DeepMind and Google Research on September 3, 2026, introduced WeatherNext 3, a global AI weather forecasting model that generates hourly forecasts at up to 5-kilometer resolution and is being integrated into Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine.

The teams described WeatherNext 3 as their most advanced and accurate global weather model to date, pointing to independent live evaluations by Brightband. A technical paper authored by researchers at Google DeepMind, Google Research, and Google accompanied the release and states that the model is running operationally.

Hourly Forecasts at Higher Resolution

WeatherNext 3 generates hourly forecasts at multiple spatial resolutions. Key surface variables such as temperature and moisture are produced at a 5-kilometer resolution, other surface variables at 10 kilometers, and atmospheric variables such as wind speed at 25 kilometers. Google said this amounts to a global weather picture roughly five times sharper than its previous model, WeatherNext 2, which produced forecasts on a 25-kilometer grid in six-hour increments.

According to the paper, WeatherNext 3 builds on the Functional Generative Network approach that underpinned WeatherNext 2 and produces 15-day, 64-member ensemble forecasts trained to minimize the continuous ranked probability score, or CRPS, a standard measure of probabilistic forecast skill. The model increases the latent size from 768 to 1024 and the mesh transformer depth from 24 to 32 layers relative to WeatherNext 2, and the production version was trained on data through June 30, 2026. The system ingests live one-hour geostationary satellite mosaics alongside traditional historical analysis, feeding a single mesh transformer that outputs dense gridded fields, discrete cyclone tracks, and station-level predictions.

Trained on Raw Observations

Most AI weather models, including WeatherNext 2, are trained on data from numerical weather prediction models, which are physics simulations that carry a six-hour data lag. Google said that lag can introduce biases for fast-changing variables such as rain and surface temperature. WeatherNext 3 instead ingests a mosaic of live global geostationary satellite data, giving it a continuously updating view of the atmosphere and allowing a new forecast every hour, each grounded in the most recent satellite observations available.

The paper states that the satellite mosaic spans 11 channels at 0.1-degree resolution with an operational latency of roughly one hour. WeatherNext 3 also trains directly on sparse weather station observations, combining three surface datasets — METAR airport stations, regional Mesonet networks, and ICOADS ship and buoy measurements — starting in 2001. The paper reports that this station output head predicts 2-meter temperature and dewpoint at any location and time, conditioned on local geographical features such as elevation.

For precipitation, the model is trained on two sources: NASA’s satellite-based Integrated Multi-satellite Retrievals for GPM (IMERG) product and the team’s own global precipitation reanalysis, called PARDIG, which the paper describes as an experimental estimate produced by a separate AI model trained to predict sparse space-borne radar data from the GPM core observatory.

WeatherNext 3 also introduces predictions aimed at renewable energy production. Google said the model forecasts 100-meter wind speeds, roughly turbine height, for wind-energy output, alongside high-resolution cloud cover and solar radiation levels to help solar farms estimate ground-level sunlight.

Reported Evaluation Results

In medium-range global precipitation forecasts, Google reported CRPS improvements of up to 60 percent against IMERG, 30 percent for the MRMS radar-based product, and 10 percent against rain gauge measurements for early lead times. The paper adds that, in evaluations against weather stations held out from training, the station head improves CRPS for 2-meter temperature by up to 30 percent compared with WeatherNext 2 and 40 percent compared with the ECMWF ENS system at short lead times.

The paper also describes a quasi-real-time evaluation over the six-week period from July 1 to August 11, 2026, using operational data feeds. In that evaluation, the authors report that WeatherNext 3 outperformed ECMWF’s AIFS ENS v2 across all upper-level variables, with an average improvement of roughly 10 percent in the first forecast week, and showed small but consistent improvements over WeatherNext 2 in ensemble mean tropical cyclone track and intensity error.

The paper documents limitations as well. The authors note visible spatial and temporal artifacts in individual forecast samples, including hexagonal patterns reflecting the model’s mesh structure, temporal discontinuities across six-hour boundaries in some output heads, and a per-member global warm-or-cold bias in the station head. They state that marginal statistics such as ensemble quantiles are largely artifact-free except in Antarctica, where station data is sparse.

Availability Across Google Products

WeatherNext 3 begins powering weather experiences in Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine starting September 3, 2026. Google said that when people plan a day or more ahead, they will see up to 50 percent more accurate precipitation forecasts, with the greatest improvements in regions where forecasts have historically been less reliable.

Researchers, developers, and businesses can query the data in BigQuery and Earth Engine or bulk-download it from Google Cloud Storage in Zarr format. Google’s developer documentation describes the model as delivering 15-day global probabilistic forecasts initialized hourly across 64 ensemble members, and lists a custom inference option on Google Cloud that lets customers generate tailored forecasts on dedicated accelerators with control over ensemble size and forecast horizons. The forecasts can also be viewed in real time through Weather Lab, Google DeepMind’s interactive visualization site.

Google’s announcement carries a disclaimer that WeatherNext is an automated, experimental AI system and that users should refer to local meteorological agencies or national weather services for official forecasts, severe weather warnings, and public safety advisories.