A research home, not a weather app
WeatherNext is now the main code home for three generations of Google DeepMind weather models. WeatherNext 2 handles global medium-range atmospheric and cyclone forecasting, while the repository keeps the older GraphCast and GenCast implementations available behind their own documentation. That consolidation is useful: researchers can compare model generations without hunting through disconnected paper supplements, and the shared utilities cover rollouts, normalization, graph components, losses, and JAX-compatible xarray handling.
The scope is also easy to misunderstand. This is not an app that answers whether it will rain tomorrow, and it is not the simplest route to putting forecast data into a product. Google already publishes daily WeatherNext 2 output through Google Cloud, WeatherLab, and Open-Meteo. If you only need predictions, use those feeds. This repository is for people who need to run, study, reproduce, or modify the models themselves.
The model choices are explained unusually well
The README does a good job of separating the available checkpoints. The operational WeatherNext 2 model works at 0.25 degree resolution, is initialized from HRES conditions, and adds 100-meter wind prediction. The cyclone checkpoints reproduce different paper and operational periods. Four model files form each full ensemble, with training cutoffs stated plainly, so researchers can avoid accidentally evaluating against data seen during training.
The Mini variants are the practical on-ramp. They operate at 1 degree resolution, use one weight file, and are intended for local tests or a single TPU or GPU. Google explicitly warns that Mini should not match the larger models. That candor matters. It prevents the convenient demo from being mistaken for evidence that the full operational system fits the same hardware or delivers the same results.
There is also a direct cyclone tracker in the demonstration workflow. The notebook walks through downloading weights, loading an initial atmospheric state, initializing the architecture, generating autoregressive forecast steps, visualizing fields, finding cyclone tracks, calculating training loss, and taking a gradient step. This is a much stronger starting point than a repository that publishes weights with only a bare inference function.
The quick start has a hardware asterisk
A first experiment is genuinely approachable. Open the linked Colab notebook, keep its default WeatherNext Cyclones Mini model, and select the free v5e-1 TPU runtime. The notebook pulls weights from Google's public cloud bucket and includes sample data paths. A P100 should also handle Mini inference. For someone evaluating the code rather than deploying it, this removes much of the usual environment work.
The installation command alone tells a different story. Installing release 0.3.0 from GitHub gives you the Python package, but not a ready forecast service, checkpoint, initial conditions, or suitable accelerator. Google recommends TPU because the implementation is optimized for it. GPU users must change the attention implementation in the notebook, and the non-Mini checkpoints require an H100 for sufficient VRAM. The larger Colab route calls for a v5p accelerator.
Training is another tier of commitment. Full training depends on ERA5 reanalysis data, preferably accessed as Zarr through WeatherBench2. Operational fine-tuning uses HRES analysis data. These are large scientific datasets with separate terms, and the README places responsibility for compliance on the user. This repository exposes training-related code, but it does not turn global weather-model training into an ordinary workstation job.
The boundaries are refreshingly explicit
WeatherNext is Apache 2.0 for code and notebooks, while other materials use CC BY 4.0. More important for an adopter, Google calls it research code supplied as-is, says it is not an officially supported product, and makes no API stability promise. The README recommends pinning a release because later changes may break callers. That is a reasonable contract for scientific software, but it should rule out treating the package as an unversioned dependency in a critical service.
The safety language is equally direct. These models were not produced with or endorsed by a government meteorological agency and do not replace official alerts. Teams may study cyclone output or build decision-support experiments, but public warning systems need authoritative sources and operational safeguards beyond this codebase.
Activity is strong, stability is still qualified
At review time, the repository had about 6,800 stars, had been pushed the previous day, and had just released version 0.3.0 with WeatherNext 2 support. GitHub reported 76 open issues and pull requests combined. Recent activity includes merged JAX fixes and documentation updates, alongside open reports about xarray changes breaking the legacy GraphCast demo notebook, dependency installation, and rollout failures on GPU and Colab TPU. That mixture shows both active maintenance and the normal fragility of a fast-moving scientific Python stack.
The right adoption pattern is narrow and deliberate. Begin with the pinned Mini notebook, verify the variables and resolution against your research question, then budget accelerator and data work before trying a full checkpoint. Keep official forecast feeds or agency data beside any experimental output. For researchers studying AI forecasting, WeatherNext is an excellent primary implementation. For application teams that merely need weather data, running it directly creates cost and risk without much benefit.