Google DeepMind and Google Research on September 3 introduced WeatherNext 3, the tech giant's most advanced global AI weather model to date. The system marks the first time a large-scale AI weather engine refreshes hourly and ingests live geostationary satellite imagery as a direct input, pushing precipitation accuracy up by as much as 60% versus its predecessor.
Hourly forecasts at 5-km resolution
WeatherNext 3 delivers 15-day global probabilistic forecasts across 64 ensemble members, initialized every hour instead of every six. Surface temperature and moisture predictions come out at up to 0.05° — roughly 5-kilometer — resolution, five times sharper than WeatherNext 2 and refreshed six times more often. Google says day-ahead rain predictions have improved by up to 50%, with the largest gains landing on high-impact convective events like flash storms.
Live satellite pipeline
The model breaks from the standard six-hour reanalysis cycle by drawing directly from raw geostationary satellite mosaics — including GOES, Meteosat and Himawari — alongside traditional analysis data. That closed-loop pipeline lets the network react in near real time to fast-moving weather signatures rather than waiting for a fresh initialization batch from national forecasting centers.
Rolling into Google's stack
WeatherNext 3 is already feeding the weather cards inside Google Search, Google Maps and the Gemini app, and is available to enterprises through the Google Maps Platform and the WeatherNext 3 API on Google Cloud. Developers can request access via the WeatherNext program, which sits alongside DeepMind's earlier releases covered on The Robotics Media and complements Google's broader physical-AI push behind Gemini 3.8 Flash.
Under the hood, WeatherNext 3 uses a functional generative diffusion model trained on decades of atmospheric reanalysis data and fine-tuned to condition on satellite observations. Google says the release is a critical step toward closing the gap with numerical weather prediction systems while running at a fraction of their compute cost.
Reporting based on coverage from Google DeepMind, Google Blog, TechCrunch, MarkTechPost and Winbuzzer.