Google releases WeatherNext 3, an AI weather model that outperforms government forecasts
Google DeepMind's new weather forecasting model achieves higher accuracy and finer resolution than traditional supercomputer-based systems and competing AI models.
What happened
Google DeepMind and Google Research released WeatherNext 3, an AI weather forecasting model that demonstrated superior performance on Operational WeatherBench against competing deep-learning models from Microsoft, Nvidia, and the European Center for Medium-Range Weather Forecasting, as well as traditional US National Weather Service forecasts. The model achieves 5 km resolution on key variables (versus 15-25 square km for earlier AI models), can now produce hourly forecasts instead of the standard prediction every six hours, and shows 60% improvement in rain forecasting compared to WeatherNext 2. It accomplishes this through a larger architecture with 2.4 times more parameters than its predecessor, training to predict specific weather station data, and incorporating raw satellite observations collected in real-time on an hourly basis rather than relying solely on government-processed datasets. Google says it will integrate WeatherNext 3 into Search, Maps, and Gemini, and make it available on cloud platforms.
Context
AI-based weather forecasting represents a fundamental shift in meteorology since the 2018 release of decades of government weather data. Traditional forecasts rely on expensive government supercomputers running mathematical physics simulations, which, while accurate, are slow and costly—barriers that have limited access to reliable forecasts in developing regions. Machine learning models can produce comparable accuracy far faster and cheaper, with direct implications for agriculture, renewable energy reliability, and economic development. The ability to ingest raw observational data rather than pre-processed datasets promises further accuracy gains but remains technically challenging; both Google and competitor WindBorne (which claims to have incorporated raw observations since late 2025) still depend on national weather datasets for global forecasting. Higher-resolution wind, rain, and cloud cover predictions stand to improve crop yields and renewable energy planning.
What's disputed
Google claims WeatherNext 3 is the first AI model to directly incorporate raw observations for high-resolution global forecasting; however, WindBorne states its WeatherMesh 6 model has been incorporating raw observations from weather balloons and other sources since late 2025. Google countered that its forecasts achieve higher resolution globally, and both companies' models still rely on national weather datasets, indicating the claim of fully independent data assimilation remains unresolved.