Google Releases WeatherNext 2 Code and Weights; Full-Model Inference Requires H100-Class Memory

Google DeepMind has expanded WeatherNext 2 from a cloud data service into a research model that can be downloaded, modified, and run independently for inference. The newly released repository includes the Functional Generative Network (FGN) architecture, normalization and autoregressive rollout tools, a cyclone tracker, training losses, and multiple sets of pretrained weights. The code is licensed under Apache 2.0, while other materials are available under CC BY 4.0. Instead of being limited to accessing existing forecasts through BigQuery, Earth Engine, or Weather Lab, researchers can now generate their own ensemble forecasts from an initial atmospheric state.

WeatherNext 2 operates on a grid with a resolution of about 0.25 degrees—roughly 30 kilometers near the equator—and advances in six-hour increments, with forecast horizons of up to 15 days. The publicly released operational version is initialized with ECMWF HRES conditions and was trained on data through 2024. In addition to temperature, pressure, and wind fields, the same model can generate cyclone tracks. FGN injects randomness into the model weights, sampling multiple mutually consistent weather scenarios from a single initial state rather than running a full conditional diffusion process as its predecessor, GenCast, does.

The hardware requirements remain substantial. According to the official documentation, non-Mini versions require H100-class GPU memory and a different attention implementation. The 1-degree Mini version intended for local testing can run on a P100 or a free TPU, but it is not expected to reproduce the full model's accuracy. Full training also requires ERA5 and HRES data, whose licensing and download costs are not included with the repository itself.

The significance of this release is that meteorological researchers can directly examine ensemble calibration, regional biases, and failure modes during extreme events, while also testing alternative initialization data or post-processing methods. Key questions now include whether third parties can reproduce the paper's results for cyclone intensity, genesis, and tracks, and whether the model retains its advantages in unseen years and across different ocean basins. Google also explicitly warns that the programming interface may introduce breaking changes and that the model is an experimental system. Disaster-preparedness decisions should therefore continue to rely on official forecasts from national meteorological agencies.