Google Research and Google DeepMind have developed a new weather model called MetNet-3. It provides high-resolution predictions up to 24 hours ahead for core weather variables such as precipitation, temperature, wind speed and direction, and dew point. MetNet-3 outperforms traditional numerical weather prediction models and is integrated into various Google products and technologies. The model uses direct observations of the atmosphere, including data from sensors like weather stations and satellites, to train and evaluate forecasts. It also employs a technique called densification, which merges data assimilation and simulation into a single pass through the neural network. MetNet-3 produces dense forecasts with a temporal resolution of 2 minutes and a spatial resolution of 1 to 4 kilometers. It delivers accurate and reliable weather information, particularly for precipitation, and is available in the contiguous United States and parts of Europe. MetNet-3’s performance is compared to other models using metrics like Continuous Ranked Probability Score (CRPS), and it consistently achieves better results. The model offers high-resolution forecasts that are both accurate and informative for users.