DeepMind's WeatherNext Cyclones Outpace Hurricane Models by Full Day

DeepMind published a paper in Nature on August 6, 2026, showing that WeatherNext Cyclones (WN-C) delivers over a full day of lead time advantage over leading operational models on track, intensity, and wind structure. Three-day forecasts match what prior systems delivered at two days. The company is open-sourcing the model weights alongside the publication.

WN-C is a single end-to-end model. Most operational systems split the task: global low-resolution models handle track, regional high-resolution models handle intensity. The split exists because intensity depends on fine-scale turbulent physics near a cyclone's core — physics that coarse global models historically missed. WN-C merges both tasks into one model trained on 20 terabytes of global atmospheric data combined with the IBTrACS archive of nearly 5,000 historical storms. The co-training signal from IBTrACS teaches the model what extreme cyclone behavior looks like at scales where global atmospheric data alone is ambiguous.

WN-C operates at 28×28 km — 100 times coarser than traditional regional models — and still achieves state-of-the-art intensity scores evaluated on cyclones from 2023–2025. Traditional regional models run at sub-kilometer to single-digit kilometer grids. A compact variant, WeatherNext 2-mini, runs at 111×111 km on a single TPU in a free Colab notebook. The researchers flag how WN-C extracts intensity signal from coarse inputs as an open question now that weights are public.

The model generates 1,000-member ensembles per cyclone, up from 50 members last season, using Functional Generative Networks that inject structured noise directly into the architecture to keep each scenario physically self-consistent. A 1,000-member, 15-day forecast runs in under one minute on a single TPU. Forecasters can surface tail risks like rapid intensification without waiting for a supercomputer queue.

The 2025 Atlantic hurricane season served as live evaluation. WN-C ran alongside the National Hurricane Center's operational workflow and predicted Hurricane Melissa's rapid intensification and Jamaica landfall far enough in advance that NHC issued an early warning the agency described as historic. CIRA Research Scientist Dr. Kate Musgrave found it showed comparable or greater skill than the best operational models for track and intensity. Partners included CIRA, the UK Met Office, the University of Tokyo, and Japan's Weathernews Inc.

The open-source release covers three variants: WeatherNext Cyclones (the version that ran during the 2025 hurricane season), WeatherNext 2 (operationalized in October 2025, surpassing the prior version on 99.9% of variables from 0–15 days, generating forecasts 8× faster), and WeatherNext 2-mini. Code and weights are on GitHub. WeatherNext 2 forecast data is available in Earth Engine and BigQuery; Vertex AI offers an early-access inference endpoint. Tropical cyclones have caused more than 700,000 deaths and $1.4 trillion in economic losses globally over the past 50 years.

For architects building weather-dependent pipelines in energy, logistics, supply chain, or agriculture: sub-minute ensemble inference on a single TPU, open weights, and BigQuery-accessible forecast data means WN-C is integrable today.