Google DeepMind releases WeatherNext 3 AI weather model with hourly satellite forecasts

Google DeepMind and Google Research have launched WeatherNext 3, an advanced AI weather model that delivers hourly global forecasts at an unprecedented 5-kilometer resolution. By training directly on live geostationary satellite imagery and sparse weather station data, the model bypasses traditional supercomputer simulation delays, reducing data lag to three to four hours. WeatherNext 3 topped the independent Operational WeatherBench leaderboard, outperforming models from Microsoft, Nvidia, and the US National Weather Service. The model is now live across Google Search, Maps, and Gemini, and features specialized outputs for wind and solar energy grid operators.

WeatherNext 3 launch

• ▪Google DeepMind and Google Research released WeatherNext 3, an advanced global artificial intelligence weather forecasting model, on September 3, 2026.
• ▪WeatherNext 3 is built on a Functional Generative Network mesh transformer architecture that scales up its predecessor's latent size from 768 to 1,024 and its mesh depth from 24 to 32 layers.
• ▪WeatherNext 3 runs a 64-member ensemble forecast cycle to generate probabilistic outputs that quantify forecast uncertainty.

Real-time satellite data training

• ▪By incorporating live satellite observations directly, WeatherNext 3 reduces typical weather forecasting data lag from roughly seven hours to three or four hours.
• ▪WeatherNext 3 ingests an 11-channel satellite mosaic of live geostationary satellite imagery directly at 0.1-degree resolution with an operational latency of approximately one hour.
• ▪WeatherNext 3 trains on sparse weather station observations from three networks, including METAR airport stations, regional Mesonet ground networks, and ICOADS ship and buoy measurements collected since 2001.

Accuracy benchmark performance

• ▪According to Google's evaluations, WeatherNext 3 achieved up to a 30 percent temperature accuracy improvement over WeatherNext 2 and up to 40 percent over the European Centre for Medium-Range Weather Forecasts' ENS system at short lead times.
• ▪WeatherNext 3's technical paper documents limitations, including hexagonal spatial patterns reflecting the transformer's mesh structure, temporal discontinuities across six-hour boundaries, and a per-member temperature bias.
• ▪WeatherNext 3 achieved the top ranking on the Operational WeatherBench, an independent forecasting leaderboard built by the startup Brightband.
• ▪WeatherNext 3 outperformed deep learning models from Microsoft, Nvidia, and the European Centre for Medium-Range Weather Forecasts, as well as traditional physics-based forecasts from the US National Weather Service.

Hourly forecast resolution improvements

• ▪WeatherNext 3 improves precipitation forecasting accuracy by up to 60 percent over WeatherNext 2, utilizing training data from NASA's IMERG dataset and Google's satellite-radar-based precipitation analysis.
• ▪WeatherNext 3 produces hourly global forecasts, a significant frequency increase compared to the standard six-hour forecast cycle used by WeatherNext 2.
• ▪WeatherNext 3 visualizes key surface variables, including temperature and moisture, at up to a 5-kilometer resolution, which is five times sharper than the 25-kilometer grid of WeatherNext 2.

Google product integration

• ▪On September 3, 2026, Google integrated WeatherNext 3 to power weather information across Google Search, Google Maps, the Gemini app, and the Google Maps Platform Weather API.
• ▪Google's official product documentation advises users to consult local meteorological agencies or national weather services for official forecasts, severe weather warnings, and public safety advisories.
• ▪WeatherNext 3 forecast data is accessible to researchers, developers, and businesses through Google BigQuery, Google Earth Engine, and Google Cloud Storage.

Renewable energy grid applications

• ▪WeatherNext 3 provides high-resolution cloud-cover and solar radiation data at the ground level to help solar farm operators estimate solar irradiance and schedule generation against grid demand.
• ▪WeatherNext 3 generates specialized forecasts for renewable energy operations, including wind speeds at 100 meters above the ground to estimate wind turbine power output.

Debatable claims

• ▪AI models should replace physics-based simulations as the primary tool for global weather forecasting
• ▪Tech platforms should prioritize official government forecasts over proprietary AI models in consumer apps