Run and verify typhoon and medium-range weather prediction models
When you need to know where a typhoon will move or how the weather will change in a few days, open this repository. Google DeepMind has released the code for the WeatherNext 2 model, allowing researchers and developers to run the model directly and visualize the results. Instead of complex formulas, you can see the actual prediction process through code and even try tracking typhoons.
WeatherNext 2 and previous models
This repository contains the code for the latest model, WeatherNext 2 (WN2), and the typhoon-specific model WeatherNext Cyclones. WN2 predicts global medium-range weather and typhoons together, calculating winds up to 100m in height. Code for the previous generation models, GraphCast and GenCast, is also provided to enable comparative experiments. Each model uses independently trained weights, but the typhoon prediction algorithm operates identically.
Model versions and resolution
The operational versions, WeatherNext2_<2025> and WeatherNextCyclones_<2025>, have a 0.25-degree resolution (approximately 30km) and were trained on data up to 2024. In particular, the Cyclones model is the version that operated in real-time during the 2025 Atlantic hurricane season. In environments with limited memory or compute resources, you can use the 1-degree resolution Mini version. The Mini version has lower performance than the larger models but is suitable for local testing or running on a single GPU/TPU.
How to run and requirements
The easiest way to start is to open the Colab Notebook. The default setting is the Mini model, which can be run on a free v5e-1 runtime. Running larger models requires a v5p accelerator. Execution is optimized for TPU, but if using a GPU, you must change the attention implementation method. Non-Mini models require an H100 for sufficient VRAM, while the Mini model can perform inference on a P100. Pre-trained weights and sample data can be downloaded from a Google Cloud Bucket.
Data and license considerations
To run the model, you must obtain the ECMWF ERA5 dataset in Zarr format, and the operational fine-tuning data uses HRES data from WeatherBench2. These data may be subject to separate terms and conditions, so you must check the restrictions before use. The code is distributed under the Apache 2.0 license, and the remaining materials under the CC BY 4.0 license. This is not an officially supported Google product and is an experimental research project that does not replace official alerts or notifications from government meteorological agencies.
