Google DeepMind described WeatherNext 3, its new global weather model, now running operationally. The model issues a fresh forecast every hour instead of every six. It does this by reading geostationary satellite images directly, which arrive far sooner than conventional data. Earlier versions relied only on processed analysis data and inherited its known biases. WeatherNext 3 also predicts rainfall estimates derived from satellites, tropical cyclone attributes and weather station readings. Its station component gives temperature and dewpoint at any point on Earth, including times between fixed steps. Errors there were lower than competing global models, including at stations held back from training. Coverage gaps in mountains and high-latitude oceans caused biases, which the team reduced with synthetic station points. One radiation variable performed poorly and is not published. The results come from a paper by the model's authors, evaluated over 2024.

• new forecast every hour, was 6-hourly
• 0.1° grid for single-level variables
• 15-day, 64-member ensemble
• satellite input latency under 1 hour