wtf is weathernext?

WeatherNext is a research project from Google DeepMind and Google Research that predicts global weather using machine learning instead of the traditional physics based simulations most weather services rely on. The repository holds the code for WeatherNext 2, the newest and most advanced model in the family, along with the code for two earlier models known as GraphCast and GenCast. The models forecast medium range weather, meaning several days out, at a global scale, and one version specifically predicts cyclone tracks and intensity. A related model, WeatherNext Cyclones, was actually used to help track storms during a real Atlantic hurricane season. The main difference between WeatherNext 2 and WeatherNext Cyclones is that WeatherNext 2 additionally forecasts wind at 100 meters above the ground. If you just want the forecast data itself rather than running the model, Google offers several ways to access it directly, including Google Cloud services like Earth Engine, BigQuery, and Vertex AI, a dedicated site called WeatherLab, and a third party service called OpenMeteo with its own API. For people who want to run the model, the easiest starting point is an interactive Colab notebook that walks through loading pretrained weights, feeding in initial weather conditions, running the forecast, and visualizing results like temperature and wind speed. The notebook defaults to a lightweight Mini model that can run on a free Colab TPU, while the larger models need more powerful hardware such as an H100 GPU or a TPU v5p. The project is written in Python and installed directly from GitHub with pip. The maintainers are explicit that this is research code provided as is, without guarantees of stability, so pinning to a specific release is recommended. Training the models from scratch requires downloading a large public weather dataset called ERA5.