DeepMind’s WeatherNext Adds a Day of Cyclone Warning and Opens Its Weights

Google DeepMind says its WeatherNext Cyclones model matches the accuracy of leading operational cyclone forecasts a full day earlier, and it released the code and trained weights under Apache 2.0.
artificial-intelligence
data-science
Author

Kabui, Charles

Published

2026-09-15

Keywords

weather-forecasting, cyclone-prediction, google-deepmind, open-weights, foundation-model

Download as PDF

Google DeepMind’s WeatherNext Cyclones predicts a tropical cyclone’s track, intensity, and wind structure about a day earlier than leading operational models. Across storms from 2023 to 2025, the team reports an average lead-time gain of at least 24 hours, an improvement it compares to a decade of conventional progress. It trains end to end on nearly 20 terabytes of atmospheric data and a database of about 5,000 historical storms. It generates 1,000-member ensembles, up from 50, each one a plausible future, and it produces a 15-day forecast in under a minute on a TPU. DeepMind published the results in Nature and released the code and weights under Apache 2.0.

Forecasters used an earlier version in the 2025 Atlantic season, when it called Hurricane Melissa’s rapid intensification and Jamaica landfall five days out, at 80% confidence and near certainty three days before. Rapid intensification, a jump of at least 35 miles per hour in 24 hours, is what makes them so dangerous. The National Hurricane Center’s verification report named it the top individual model for both track and intensity that season. Open weights mean weather services and researchers can run it without a vendor contract, and a miniature version fits a free Colab notebook.

The result also challenges a long-standing assumption. WeatherNext reads coarser input than the regional models it outperforms, evidence that high resolution is not required for accurate intensity forecasts and that low-resolution data carries more storm signal than meteorologists recognized.

Read More: Groundsource: Google Turns 25 Years of News into 2.6 Million Flood Records.

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {DeepMind’s {WeatherNext} {Adds} a {Day} of {Cyclone}
    {Warning} and {Opens} {Its} {Weights}},
  date = {2026-09-15},
  url = {https://toknow.ai/posts/weathernext-cyclones-extra-day-warning-open-weights/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “DeepMind’s WeatherNext Adds a Day of Cyclone Warning and Opens Its Weights.” https://toknow.ai/posts/weathernext-cyclones-extra-day-warning-open-weights/.