Google DeepMind published WeatherNext Cyclones in Nature on 6 August, reporting more than a full day of additional lead time over leading operational forecasts for cyclone tracks, intensity and wind structure. Three-day forecasts now match the accuracy prior models reached at two days.
The error bars
Track error at the three-day window is approximately 100 km; intensity error is around 11 knots. The model was evaluated on cyclones from 2023-2024 — some secondary coverage says 2023-2025, which DeepMind's own post does not support.
What it was trained on
Nearly 20 terabytes of global atmospheric data plus IBTrACS, the historical best-track archive covering nearly 5,000 storms. It was built with the US National Hurricane Center, the Cooperative Institute for Research in the Atmosphere and the UK Met Office. WeatherNext 2 and WeatherNext Cyclones are on GitHub with code and model weights, alongside a WeatherNext 2-mini.
Where the gain is, and is not
The headline advance is on track. Intensity forecasting — whether a storm rapidly strengthens before landfall — remains the harder problem and the one that most often determines damage. An 11-knot error is progress, not a solved case.
Not a replacement
It operates alongside National Hurricane Center forecasters rather than replacing official warnings. DeepMind's line that the result represents a decade's worth of meteorological progress is the company's own analogy, not a metric in the paper — which is peer-reviewed, unlike most model claims that reach this level of coverage. The release of code and weights alongside it means the comparison can be rerun by the agencies that would have to trust it, which is the part that decides whether it gets used operationally.