# How DeepMind's WeatherNext Is Changing Cyclone Forecasting > Source: > Published: 2026-08-30 08:40:04+00:00 Tropical cyclones are among the deadliest and costliest natural disasters on Earth, responsible for over 700,000 deaths and an estimated $1.4 trillion in economic damage worldwide over the past five decades. For meteorologists, every extra hour of accurate warning can mean the difference between an orderly evacuation and a catastrophe. On August 6, 2026, Google DeepMind published research in *Nature* introducing **WeatherNext Cyclones**, an AI model that pushes the boundary of how far in advance a storm's path, strength, and wind structure can be reliably predicted and then open-sourced the entire system for researchers worldwide. Cyclone prediction has traditionally forced meteorologists to choose between two separate modeling approaches. A storm's **track** is driven by large-scale atmospheric currents, which coarse, global-scale models capture well. A storm's **intensity**, however, depends on fine-grained thermodynamic processes near the storm's core, which has historically required specialized, high-resolution local models. Running both in parallel is expensive and doesn't always produce a consistent picture of the storm. WeatherNext Cyclones (WN-C) is a single model that predicts track, intensity, and wind structure together, in one autoregressive system, rather than stitching together outputs from separate models. It was trained end-to-end on nearly 20 terabytes of global atmospheric data alongside the IBTrACS historical database, which covers close to 5,000 past storms. At the core of the system is a technique called **Functional Generative Networks (FGNs)**, which replace the diffusion-based approach DeepMind used in its earlier GenCast model. FGNs let WeatherNext generate large ensembles of possible storm outcomes cheaply up to 1,000 simulated scenarios per forecast which helps forecasters gauge the probability of rare but dangerous events like sudden rapid intensification. Despite operating on a resolution roughly a hundred times coarser than conventional physics-based systems, the model still matches or beats them on accuracy a result DeepMind's own researchers describe as not yet fully understood. The efficiency gains are notable too: a full 15-day forecast can be generated in under a minute on a single tensor processing unit (TPU). Benchmarked against historical cyclones from 2023 through 2025, WeatherNext Cyclones delivered roughly a full extra day (24 hours) of reliable warning compared with leading operational systems. A 3-day WeatherNext forecast was found to be about as accurate as what earlier systems achieved at 2 days. At the 5-day mark specifically, the model's average track error was around 230 kilometers, compared to roughly 370 kilometers for the ECMWF ensemble system and 335 kilometers for DeepMind's own earlier GenCast model. In simulated blends with official consensus forecasts, WN-C improved track accuracy by about 28 percent and intensity accuracy by about 6 percent a jump DeepMind's team compared to roughly a decade's worth of typical progress in traditional forecasting. WeatherNext wasn't just a benchmarking exercise. During the 2025 Atlantic hurricane season, the U.S. National Hurricane Center used the model operationally to help forecast Hurricane Melissa's rapid intensification and its landfall in Jamaica an early sign that AI-based forecasting tools are moving from research demos into real forecasting workflows. DeepMind released the code and pretrained weights for three variants WeatherNext 2, WeatherNext Cyclones, and a lightweight WeatherNext 2-mini that can run in a free Google Colab notebook under an Apache 2.0 license on GitHub. Live forecasts are also available through the Weather Lab platform inside Google Earth AI. The work was developed in collaboration with the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere (CIRA), and the UK Met Office. Importantly, DeepMind frames WeatherNext as a tool to support forecasters, not replace them official storm warnings still come from national weather services. WeatherNext is a good example of where AI is currently having its most tangible impact in science: not in speculative future capability, but in incrementally compounding gains on a well-defined, high-stakes physical prediction problem. An extra 24 hours of cyclone warning is not a marginal number it is measured in evacuation windows, supply chain decisions, and ultimately lives. As the model's code and weights are now public, WeatherNext also sets a precedent for how AI weather research can be shared openly rather than locked behind proprietary forecasting products. **Sources**