For cyclone forecasting, an extra day can change evacuation planning, port operations, and grid preparation. Google DeepMind said in an August 6 X post that its WeatherNext model reaches state-of-the-art accuracy for predicting storm track and intensity, giving forecasters an average of 24 additional hours to prepare.

Google DeepMind described the gain as “a critical extra 24 hours.”

The work is published in Nature, and the linked DeepMind blog says WeatherNext enables more accurate cyclone forecasts while making the model openly available. A follow-up post in the same thread adds that the code and model weights are being released on GitHub for academic work, operational forecasting, and more specialized local models.

Google DeepMind’s account is the lab’s direct channel for research releases across general AI and scientific systems such as AlphaFold, AlphaEarth, and WeatherNext. This post fits the latter pattern: the claim is not a chatbot feature or an API tier, but a domain model aimed at a measurable public-safety problem. Cyclone forecasting depends on both track and intensity, and errors in either one can change who evacuates, where resources are staged, and when emergency warnings go out.

The open release matters because weather models need more than headline accuracy. Meteorological agencies will want to test regional bias, rare storm behavior, data latency, and how WeatherNext complements existing numerical weather prediction systems. Open weights and code give outside researchers a better path to reproduce the result, stress-test it, and adapt it for local forecasting conditions.

The next things to watch are the Nature paper’s detailed metrics, the GitHub license and reproducibility notes, and whether national weather services or disaster-response agencies begin piloting WeatherNext alongside their current forecast stack.

Source: Google DeepMind on X

#weather-ai

WeatherNext adds 24 hours to cyclone forecasts with open weights