Google DeepMind unveiled WeatherNext Cyclones on August 6, 2026, an AI model that forecasts the track, intensity and wind structure of tropical cyclones up to 15 days ahead. The company says the model adds about a day of accuracy to three-day forecasts, roughly a decade of progress in meteorology.
Key takeaways
- Unveiled August 6, 2026, the WeatherNext Cyclones model
- Forecasts cyclone track, intensity and wind structure up to 15 days ahead
- About one extra day of accuracy on three-day forecasts
- A forecast is produced in under a minute on a TPU, with an ensemble of up to 1,000 members
- The model code and weights were released as open source
Faster and further
Google DeepMind positions WeatherNext Cyclones as a model reaching state-of-the-art accuracy across three dimensions at once: track, intensity and wind extent. Track forecasts were compared with the ECMWF ensemble model?ensemble model: A forecasting method that runs many simulation variants at once to estimate uncertainty, instead of a single point forecast. and intensity with the HWRF system, on cyclone data from 2023 and 2024.
Speed matters as much as accuracy. A single forecast is produced in under a minute on one TPU, and the model scales to a 1,000-member ensemble, which helps estimate uncertainty. Classic physics-based models need far more compute and time by comparison, and running such a large ensemble is often beyond their reach. More members mean a sharper picture of the possible tracks and intensity of a cyclone, not a single point forecast.
An open model and its use
Google DeepMind released the model code and weights?weights: A model's trained parameters. Releasing them lets anyone run the model themselves, without training from scratch. as open source, and a smaller WeatherNext 2-mini variant runs in a free Colab notebook. The model was built with institutions including the US National Hurricane Center, the UK Met Office and CIRA. An earlier version already supported NHC forecasts during Hurricane Melissa in 2025.
Why it matters
Cyclone forecasting is one of those areas where an extra day of lead time genuinely saves lives and property, it gives more time to evacuate and prepare. AI models change the economics of forecasting here: instead of hours on a supercomputer, a minute on a single chip is enough, which makes it cheap to run hundreds of variants and better measure uncertainty. Releasing the code and weights as open source raises the chance that agencies and centers outside the largest budgets can use the tool too. That moves advanced weather forecasting toward widely available infrastructure.
What's next
- The model and weights are publicly available through GitHub and the Weather Lab interface, the next test is operational use by weather services in the 2026 cyclone season
- Collaboration with NHC, the Met Office and NOAA points to validation on real events, not just historical data
Sources
- Google DeepMind, WeatherNext AI model achieves breakthrough in forecasting cyclones
