Google DeepMind and Google Research today introduced WeatherNext 3, a new artificial intelligence model designed to produce more frequent and higher-resolution global weather forecasts. Google plans to incorporate outputs from the model into Search, Google Maps and Gemini, and to make WeatherNext 3 available to users and researchers via Google Cloud.
Capabilities of WeatherNext 3
WeatherNext 3 provides forecasts for key variables at a spatial resolution of 5 km and can produce hourly predictions rather than the more common six-hour cadence. Google reports roughly a 60% improvement in rainfall prediction compared with WeatherNext 2. The new model is larger than its predecessor, with about 2.4 times more parameters, and the development team adjusted the decoder heads’ targets to generate more useful outputs.
Performance versus other systems
On Operational WeatherBench — a benchmarking utility for AI forecasts developed by the startup Brightband — WeatherNext 3 ranked as the most accurate among the models tested. The benchmark evaluates metrics such as temperature, wind speed and humidity. According to Google, the model outperformed other deep-learning systems from Microsoft, Nvidia and the European Centre for Medium-Range Weather Forecasts (ECMWF), and also beat traditional forecasts from the U.S. National Weather Service and the ECMWF.
Design choices behind the improvements
The improvements stem from several specific design choices. WeatherNext 3 can ingest raw, real-time observational data — including hourly satellite inputs — which enables more frequent updates. The developers also trained the model to target forecasts at specific weather stations; this helps both to provide finer spatial detail and to enable direct evaluation against ground-truth station measurements.
Daniel Rothenberg, an atmospheric scientist at Brightband, noted that the ability to predict what a particular site (for example, a given airport’s weather station) will record on an hourly basis brings the forecasting task closer to observable measurements.
Data sources and remaining limits
Google described WeatherNext 3 as the first AI model to directly incorporate raw observations for a high-resolution global forecast. The startup WindBorne, however, says its model WeatherMesh 6 has been incorporating raw observations from its fleet of weather balloons and other sources since late 2025. Google responded that WeatherNext 3 offers higher resolution globally. Both approaches still rely on national weather datasets for their forecasts, so fully direct data assimilation at all scales remains a work in progress.
Wider implications
Faster, lower-cost AI-driven forecasts could have important economic and social effects, particularly in lower-income regions where the cost of sensors and supercomputing has limited access to high-quality forecasts. Bill Gates has pointed to AI-powered weather forecasting as a notable societal benefit, for example by improving crop yields in developing countries. Ferran Alet, a staff research scientist manager at DeepMind, said higher-resolution forecasts of wind, precipitation and cloud cover could also make renewable energy projects more reliable.
Availability and rollout
Google says WeatherNext 3 capabilities will be integrated into Google products and made available on Google Cloud, but it has not published a detailed timetable for rollout or exact release dates.
Summary
WeatherNext 3 represents a step forward in AI-based meteorology with finer spatial resolution, hourly updates and substantially improved rainfall prediction. Together with advances from other research groups and startups, it may expand access to faster, more precise weather information across more regions.
