WeatherNext gains a day on cyclone forecasts

Google DeepMind’s WeatherNext Cyclones gained more than 24 hours of forecast lead time across storm track, intensity, and wind structure, making its three-day forecasts as accurate as previous models’ two-day forecasts. The peer-reviewed Nature paper evaluated storms from 2023 through 2025, while DeepMind’s release says the open model can generate a 1,000-member, 15-day ensemble and offers a mini version that runs on one TPU. Builders working on scientific AI systems whose value depends on expert adoption should file this as a deployment pattern: publish the weights, preserve human forecasters as decision-makers, and measure progress in extra warning time rather than benchmark points.