Google Unveils WeatherNext 3, an Advanced AI Model Enhancing Weather Forecast Accuracy

Google DeepMind and Google Research have launched WeatherNext 3, an advanced artificial intelligence model designed to improve the accuracy and frequency of weather forecasts. This model represents a significant advancement in meteorology by leveraging deep learning techniques to better understand and predict atmospheric behavior. WeatherNext 3 will integrate into various Google services, including Search, Google Maps, and Gemini, and will also be accessible to researchers and users through Google's cloud platforms. According to Samier Merchant, a senior staff engineer at Google, this is the first time core weather variables from an AI model will directly support multiple Google products. In testing on Operational WeatherBench, a benchmarking tool developed by the startup Brightband, WeatherNext 3 outperformed other AI models from Google, Microsoft, Nvidia, and the European Center for Medium-Range Weather Forecasting (ECMWF). It also surpassed traditional forecasts from the US National Weather Service and ECMWF in accuracy for key metrics such as temperature, wind speed, and humidity. Traditional weather forecasts rely on government supercomputers that process complex physical equations to simulate weather patterns. While effective, these systems are costly and relatively slow. Since the ECMWF released decades of weather data in 2018, AI researchers have developed models capable of faster predictions with comparable accuracy. WeatherNext 3 addresses common limitations of previous AI forecasting models by increasing spatial resolution to 5 kilometers, improving rainfall prediction accuracy by 60% compared to its predecessor, and providing hourly forecasts instead of updates every six hours. These improvements stem from a larger model architecture with 2.4 times more parameters and refined targeting of forecast outputs. A notable feature of WeatherNext 3 is its ability to generate predictions tailored to specific weather stations, enabling more granular forecasts and direct comparison with ground-truth data. This approach enhances the model's practical utility, as highlighted by atmospheric scientist Daniel Rothenberg from Brightband. The model's capacity for more frequent forecasting is supported by its use of real-time satellite data collected hourly. Incorporating raw observational data rather than processed outputs from supercomputers presents technical challenges but offers potential gains in forecast accuracy. Google claims WeatherNext 3 is the first AI model to integrate raw observations for a high-resolution global forecast. However, the AI weather startup WindBorne notes that its WeatherMesh 6 model has been using raw data from weather balloons and other sources since late 2025. Both models still depend on national weather datasets, indicating ongoing work is needed for fully direct data assimilation. The adoption of transformer-based AI models in meteorology is gaining traction among European and US weather agencies, promising faster and more cost-effective forecasts. This progress could have significant economic benefits, especially in regions where expensive sensors and supercomputers limit access to accurate weather information. Industry figures like Bill Gates have emphasized the importance of AI-enhanced weather forecasting for improving agricultural productivity in developing countries. Additionally, higher-resolution forecasts of wind, rain, and cloud cover can support the reliability of renewable energy projects. Ferran Alet, a staff research scientist at DeepMind, emphasized that Google's goal is to provide useful weather information to users, reflecting the widespread interest in weather conditions worldwide.