Google DeepMind and Google Research have launched WeatherNext 3, a new AI weather forecasting model that will feed into Google Search, Maps and Gemini, as well as being made available to researchers via Google's cloud platforms. The model has topped Operational WeatherBench, an independent benchmark for AI forecasts, beating rival deep-learning systems from Microsoft, Nvidia and the ECMWF, as well as traditional forecasts from national weather agencies, marking a further step in the shift from physics-based supercomputer modelling towards machine learning approaches.

The new model can forecast key variables to a resolution of 5km, compared with the 15-25km range typical of earlier AI systems, and produces hourly updates rather than the standard six-hourly forecasts. Its rainfall predictions are said to be 60% more accurate than its predecessor, WeatherNext 2, thanks in part to having 2.4 times more parameters and redesigned outputs. The model has also been trained to target forecasts at specific weather stations, such as individual airports, allowing for more precise local predictions and easier validation against real-world measurements.