Google DeepMind's WeatherNext Pushes Cyclone Forecasts a Full Day Further Out

Google DeepMind’s WeatherNext models now produce three-day tropical cyclone forecasts — track, intensity, and wind structure — as accurate as what the prior generation of models could only manage two days out. DeepMind, publishing Nature-backed research on August 6, 2026, describes that as roughly a decade of typical meteorological progress compressed into one advance. The system generates 1,000 possible forecast scenarios per storm using what DeepMind calls Functional Generative Networks, up from about 50 scenarios in the prior generation — the added scenario density is what catches low-probability events like sudden rapid intensification. A full 15-day forecast runs in under a minute on a single TPU, at 28x28km resolution, roughly 100 times coarser than traditional intensity-prediction models, without giving up accuracy. DeepMind points to the 2025 hurricane season as validation: the model helped the U.S. National Hurricane Center anticipate Hurricane Melissa’s rapid intensification and landfall in Jamaica, buying forecasters more warning time. DeepMind released the code and weights for three variants — WeatherNext Cyclones, WeatherNext 2, and a lighter WeatherNext 2-mini — under an Apache 2.0 license.

The open-source decision lands next to a related move: ECMWF’s own AI weather model, the Artificial Intelligence Forecasting System, went operational February 25, 2025 and became, per a Hugging Face write-up published July 28, 2026, the first fully operational open machine-learning weather model at its parameter range. AIFS runs on roughly 1,000 times less energy than physics-based forecasting and now has a compatibility patch letting it run on any GPU — or a CPU — instead of requiring the Ampere-class hardware its original flash-attention dependency demanded. Between DeepMind’s Apache-licensed weights and ECMWF’s CC BY 4.0 release, two organizations that used to guard forecasting models as proprietary infrastructure are now both giving them away.

For companies weighing where AI actually earns its keep outside of chat interfaces, cyclone forecasting is a clean example of the pattern: a narrow, well-defined prediction problem, decades of labeled training data, and a clear cost of being wrong. WeatherNext’s edge isn’t a bigger model — it’s more scenarios per forecast, run cheaply enough to do routinely. That’s the shape of an outcome, not a demo.