DeepMind's WeatherNext beats existing cyclone forecasts by a day, and the model is open

AI weather forecasting just bought coastal cities an extra day, and Google gave the model away.

What the Nature paper shows

In research published in Nature on August 6, 2026, Google DeepMind showed its WeatherNext model predicting tropical cyclone track, intensity and wind structure more accurately than existing systems. In DeepMind’s own words, the model delivers an extra day’s worth of predictive accuracy: three-day forecasts as good as what prior models could provide for only the next two days.

DeepMind puts that jump at roughly a decade of normal meteorological progress. Forecast skill in this field typically advances by hours per decade, not days per paper, which is what makes the claim notable enough for Nature.

Open weights, and a real-world record

The code and model weights are published on GitHub, WeatherNext 2 alongside the cyclone models, including a lightweight version the repository recommends running on a runtime that is free in Colab. The notebooks are Apache 2.0.

The model also has operational history rather than just benchmark results. During the 2025 hurricane season, the National Hurricane Center used it in forecasting Hurricane Melissa’s rapid intensification and landfall in Jamaica. Rapid intensification is precisely the failure mode that has embarrassed traditional forecasting models, which makes that deployment a meaningful stress test rather than a courtesy citation.

Why an extra day is the right unit

An extra day is not a benchmark score. It is the margin between an evacuation order that arrives in time and one that does not, between boarding up and driving out. For emergency managers, forecast lead time converts directly into decisions, and a three-day forecast with two-day quality moves every one of those decisions earlier.

The open release compounds the effect. Cyclone risk concentrates in countries that cannot buy proprietary forecasting systems, and Apache-licensed weights that run on a free Colab tier put the capability within reach of any national meteorological service with an internet connection. Among the arguments for open-sourcing frontier models, a hurricane forecast that reaches Jamaica a day earlier is about as concrete as they come.