Google WeatherNext 3 Sharpens AI Forecasting

Google is moving artificial intelligence further into mainstream weather forecasting with WeatherNext 3, a new model designed to deliver more frequent and locally detailed predictions. The system will begin feeding weather information across Search, Maps and Gemini, bringing machine-learning forecasts directly into products used by millions.

Developed by Google DeepMind and Google Research, WeatherNext 3 produces hourly forecasts for key variables at resolutions of around five kilometres. Its rainfall evaluations improved by 60 per cent over WeatherNext 2, while the model carries 2.4 times more parameters than its predecessor. Independent tests by Brightband’s Operational WeatherBench placed it ahead of rival AI systems from Microsoft, Nvidia and the European Centre for Medium-Range Weather Forecasts across measures including temperature, wind speed and humidity.

The technical shift is equally significant. WeatherNext 3 can ingest hourly satellite observations directly, reducing its dependence on processed datasets generated by traditional forecasting systems. Conventional models still rely heavily on large supercomputers running physics-based calculations, which remain accurate but expensive and comparatively slow.

AI weather forecasting is moving rapidly from research into infrastructure. European and US weather agencies are already incorporating machine-learning models into operational products, while Google’s earlier WeatherNext work has been used to improve cyclone forecasting. In August, DeepMind said its cyclone model delivered roughly an extra day of predictive accuracy and could generate a 15-day forecast in under a minute on a TPU.

WeatherNext 3 therefore represents more than a consumer weather upgrade. Faster, cheaper forecasting could improve planning across renewable energy, agriculture, aviation and disaster response. The real test is whether those gains hold when conditions are most volatile, where reliability matters more than model speed.