Google DeepMind released WeatherNext 3, a global weather-forecasting AI model that produces 5-kilometer-resolution forecasts refreshed every hour, trained on real-time weather station observations3,5,6.
The model represents a five-fold increase in sharpness over Google's previous weather model, according to the company. Google described WeatherNext 3 as capable of forecasts at "unprecedented resolution".
"One of the main developments is for [WeatherNext 3] to go beyond what data most global AI models train on," said Samier Merchant, a research engineer at Google Research. The model learns from real-time weather observations rather than relying solely on the reanalysis datasets that most global AI weather models use.
Google said it will start feeding WeatherNext 3 into weather information that users see across Google Search, Google Maps, and Gemini2. The integration marks a direct consumer-facing deployment for a model that originated in Google DeepMind's research pipeline.
ANALYSIS Routing WeatherNext 3 into Search, Maps, and Gemini simultaneously turns a research artifact into infrastructure serving billions of queries, giving Google a differentiated data layer that competing large-language-model assistants would need to source externally.
The release continues a broader shift in meteorology driven by deep learning techniques. WeatherNext 3's hourly update cadence and 5 km grid represent a new resolution tier for Google's AI weather modeling efforts4.
ANALYSIS Training directly on station observations, rather than on gridded reanalysis products that smooth raw data, could reduce the compounding of upstream errors that has limited earlier AI weather models. The practical test will be whether the 5 km resolution holds up in regions with sparse observation networks.