Google launches WeatherNext 3, more accurate AI weather forecasts

The release of WeatherNext 3 marks a step in the ongoing shift from traditional physics-based forecasting to deep learning. The ECMWF released over half a century of weather data in 2018, which allowed researchers to train models that make predictions faster and with comparable accuracy to government tools. WeatherNext 3 addresses key weaknesses of earlier AI models, such as wide forecast areas and poor rain prediction, by aiming for higher resolution and incorporating raw data.

These improvements come from design choices, including a model with 2.4 times more parameters than its predecessor and training that targets specific weather data stations. This is significant because it allows for more granular predictions and evaluation against ground-truth data. However, the model still relies on national weather datasets, and a startup named WindBorne says its model has incorporated raw observations since late 2025, suggesting this area is still developing. The broader implication is the potential for improved forecasts to benefit sectors like renewable energy and agriculture in developing countries, as cited by Bill Gates.