Google has open-sourced its WeatherNext AI weather prediction model, releasing both the code and the model's trained data on GitHub. Developed by researchers at Google DeepMind and Google Research, the model can produce 15-day forecasts of a cyclone's path and strength, with the goal of giving people earlier hurricane warnings.

The release comes roughly a year after Google introduced the second generation of WeatherNext. The results have now been published in Nature, one of the world's top scientific journals, which means the model's performance has been checked by independent scientists rather than just claimed in a company blog post.

To train the model, researchers used nearly 20 terabytes of global atmospheric data, plus historical cyclone records from an international archive called IBTrACS. Think of it as giving the AI several decades of recorded hurricane behavior to study, so it can learn the patterns of how storms develop and move over time. When it comes to actually making a prediction, WeatherNext generates a full 15-day forecast in under a minute on a TPU, which is Google's custom chip for running AI. Traditional weather forecasting relies on simulating the physics of the atmosphere step by step, which requires far more computing power and time.

Google did not build this alone. The National Hurricane Center, the Cooperative Institute for Research in the Atmosphere (CIRA), the UK Met Office, and other weather agencies all contributed. That collaboration matters because the agencies that helped build the model are the same ones responsible for issuing public weather warnings. Having them involved during development means the model can move more quickly from a research project to something used in real emergency situations.

The decision to release both the code and the trained model is worth noting. In recent years, several AI weather models — including Google's earlier GraphCast, plus offerings from Nvidia and Huawei — have been released with varying degrees of completeness. Some shared the trained model but not enough code for others to fully reproduce the work. By releasing everything, Google makes it easier for weather services and university researchers around the world to test, adjust, and use the model in their own forecasting systems. It also lets outside experts independently verify how well the model actually performs, which has been a point of tension as AI-based weather forecasting grows up.

The 15-day forecasting window for cyclone path and strength is where WeatherNext could make the biggest real-world difference. Lead time — how far in advance you can warn people — is the most important factor in hurricane preparedness. Even small improvements in forecast accuracy at longer ranges mean smaller evacuation zones and better placement of emergency supplies. If the cyclone predictions hold up under independent testing, the model could supplement existing forecasting systems rather than replace them. Weather services have always layered new methods on top of old ones, combining different models to account for uncertainty. An AI model that produces forecasts in under a minute and looks two weeks ahead fits naturally into that approach as one more tool in the toolkit.

The broader context here is how quickly large-scale AI and atmospheric science are converging. A few years ago, using AI for weather prediction was a research curiosity. Now these models are matching or beating traditional methods on standard accuracy measures, and at a fraction of the computing cost. Google's decision to make WeatherNext freely available speeds that trend along, because any organization with enough computing power can run it. The agencies that helped build it will likely be the first to test it in real forecasting operations, but the open release means the wider research community can dig into its strengths, weaknesses, and failure modes at the same time.

For anyone who has watched the tech industry over the past few years, the pattern is familiar: a powerful AI model, trained on a large dataset, achieves strong results, gets peer-reviewed, and is then released openly to speed up adoption and scrutiny. What is less familiar is the setting. Weather prediction is not a phone app or a software tool — it is a public-safety function where the stakes include human lives. The speed at which AI-based forecasting is moving from research papers to real-world use is, by any measure, remarkably fast.