WeatherNext: AI model achieves breakthrough in forecasting cyclones

Forecasting a hurricane has always meant juggling two different jobs. One model tracks where the storm is headed, using coarse global weather patterns. Another tries to nail how strong it'll get, which depends on messy, small-scale physics near the storm's core. Google DeepMind's new WeatherNext system, detailed in a Nature paper published August 6, 2026, does both with a single AI model — and does it better than the specialized tools meteorologists have relied on for decades.

The numbers are the headline here. WeatherNext Cyclones gives forecasters roughly 24 extra hours of reliable lead time on track, intensity, and wind structure. A three-day forecast now carries the accuracy that used to require two days of runway. DeepMind frames that jump as equivalent to ten years of normal meteorological progress, based on trends measured over the past two decades. For an event that historically kills people by the hundreds of thousands and costs over $1.4 trillion globally since the 1970s, a single extra day of warning is not a small thing.

What's strange is how the model gets there. Conventional wisdom said intensity forecasting demanded very high-resolution data to capture the fine thermodynamic churn inside a storm. WeatherNext Cyclones instead runs on 28x28km grids, a hundred times coarser than what specialized models typically need, and a smaller variant called WeatherNext 2-mini works at an even blunter 111x111km resolution while still performing well. DeepMind admits it doesn't fully understand why this works, which is a refreshingly honest admission for a company usually selling certainty.

The model was trained end-to-end on nearly 20 terabytes of atmospheric data plus the IBTrACS record of almost 5,000 historical storms, then paired with something called Functional Generative Networks to spit out ensemble forecasts fast — a full 15-day, 1,000-scenario run in under a minute on a TPU. That ensemble size matters practically. During the 2025 season, the system helped the National Hurricane Center anticipate Hurricane Melissa's rapid intensification before it hit Jamaica, giving crews on the ground actual extra hours to prepare rather than theoretical ones.

Now DeepMind is releasing the code and weights for WeatherNext 2, WeatherNext Cyclones, and the mini version, which can run in a free Colab notebook on a single TPU. That's a real departure from keeping this kind of infrastructure locked inside a research paper or a paid API. Whether local forecasting agencies with limited compute budgets can actually put these models into daily operational use, alongside the human forecasters who still have to make the final call, is the part worth watching over the next few storm seasons.