Google’s latest AI weather model gives you no excuse to forget your umbrella

Google DeepMind and Google Research unveiled WeatherNext 3, an artificial‑intelligence model that dramatically sharpens weather forecasting by delivering high‑resolution, hourly predictions across the globe. The system, announced today, will be integrated into Google’s consumer products—including Search, Maps and the Gemini AI assistant—and made available to external users via Google Cloud. According to senior staff engineer Samier Merchant, WeatherNext 3 will become the first AI model whose core variables directly power multiple Google services. In benchmark tests on Operational WeatherBench—a comparative platform created by the startup Brightband—the model outperformed competing deep‑learning systems from Google, Microsoft, Nvidia and the European Centre for Medium‑Range Weather Forecasts (ECMWF), as well as traditional forecasts from the U.S. National Weather Service and ECMWF itself. It achieved superior accuracy on key metrics such as temperature, wind speed and humidity, marking a significant step forward in AI‑driven meteorology.

The breakthrough stems from addressing three long‑standing weaknesses of earlier AI forecasts: coarse spatial resolution, limited rain prediction, and reliance on pre‑processed government datasets. WeatherNext 3 now predicts weather at a 5‑kilometre grid, a five‑to‑sixfold improvement over its predecessor’s 15‑to‑25 km coverage, and its rain forecasts are 60 percent more accurate than those of WeatherNext 2. Crucially, the model can generate hourly forecasts rather than the conventional six‑hour intervals, thanks to its larger architecture—2.4 times more parameters—and a redesign of decoder heads that yields more useful outputs. Researchers also trained the system to target specific weather stations, enabling direct comparison with ground‑truth observations and improving granularity for locations such as Denver’s airport. By ingesting raw satellite data in real time, WeatherNext 3 sidesteps the slower, computationally heavy supercomputer pipelines that traditionally dominate forecasting, though it still depends on national weather datasets for baseline inputs.

The deployment of WeatherNext 3 promises broader economic and societal impacts, especially for regions where expensive supercomputers and dense sensor networks are unaffordable. Faster, cheaper, and more precise forecasts can enhance agricultural planning, disaster preparedness and daily decision‑making for millions of users who will see the improved predictions in Google’s consumer interfaces. The model joins a growing suite of AI tools already adopted by European and U.S. weather agencies, and its success reinforces the transformer‑based revolution that runs parallel to the rise of large language models. Industry leaders such as Bill Gates have highlighted AI‑enhanced weather forecasting as a key lever for boosting crop yields in developing nations. While competitors like WindBorne’s WeatherMesh 6 claim earlier use of raw observational data, Google emphasizes WeatherNext 3’s superior global resolution. Nonetheless, both systems still rely on national datasets, indicating that further advances in direct data assimilation will be needed before AI can fully replace traditional supercomputer‑driven models.

⚡ Effects Interpreter

🌍 Effects for your country