DeepMind has released WeatherNext 3, a global ensemble AI weather model that marks a significant architectural shift by moving from ERA5 reanalysis data to direct ingestion of raw satellite imagery (infrared, visible, and water vapor channels) and station METAR data. Unlike traditional Numerical Weather Prediction (NWP) models that rely on complex data-assimilation pipelines and physics equations, WeatherNext 3 utilizes a Graph Neural Network (GNN) backbone with a learned mesh representation and a direct observation encoder to bypass heavy pre-processing. This enables an hourly forecast cadence, a major improvement over the 3–6 hour cycles of standard NWP models. Key technical advancements include a 5 km resolution for temperature and humidity—a 5–6x linear improvement over the 28 km resolution of previous-generation models like GraphCast—and 10 km resolution for wind and pressure. While the model excels at precipitation nowcasting and rapid updates for renewable energy and aviation, independent analysis from the German Meteorological Service (DWD) highlights potential risks, including a lack of explicit measurement error modeling, potential physical inconsistency regarding mass and energy conservation, and the inability to extrapolate to unprecedented climate regimes not present in the training data. Currently, the model is integrated into Google Search, Maps, and Gemini.