WeatherNext 3 on GitHub: Repos, Open Models and Data Access
Searching for WeatherNext 3 code leads to the right family home: google-deepmind/weathernext. What lives there today needs one clarification - the repository currently packages the WeatherNext 2 generation, the Cyclones line used in operational forecasting, and the older GraphCast and GenCast models, while WeatherNext 3 itself is documented through its paper and served through cloud data channels. This page maps what is open, what is hosted, and the shortest path from a repository to real forecast data. For the request-level route instead, read API and data access.
7-day free trial, then $19/mo · cancel anytime · compare plans
The repositories that matter
One organization, a small set of repositories, each with a different job.
google-deepmind/weathernext
The family home: code for WeatherNext 2's forecasting model, the WeatherNext Cyclones models, shared utilities, notebooks and documentation in one place.
WeatherNext 2 code and weights
Pretrained checkpoints for the operational line are published as npz weight files, with an interactive Colab notebook that runs the model end to end.
WeatherNext Cyclones
The tropical cyclone models that ran operationally during the 2025 Atlantic season, including lightweight Mini variants that fit smaller accelerators.
GraphCast and GenCast
The older generation deterministic and diffusion-based models, still hosted in the same repository for research and comparison.
Open code, open weights, hosted data - which is which
The WeatherNext family mixes licensing and delivery models. Reading the map correctly avoids the two classic assumptions: that everything is open, or that everything needs a GPU.
Code and notebooks: Apache 2.0
The repository's code and Colab notebooks are licensed under Apache 2.0, with other materials such as weights and documentation under CC BY 4.0. Check the terms before production use.
Weights you can run yourself
Pretrained checkpoints live in a Google Cloud bucket. The Mini variants run on modest accelerators; the full models want serious hardware such as H100-class GPUs or TPUs.
Hosted data if you skip the model
Forecast outputs are distributed through Earth Engine, BigQuery and Cloud Storage, and via services such as Open-Meteo and Weather Lab, without running any model at all.
The WeatherNext 3 paper
The current generation is documented through its technical report (arXiv 2609.03582) and Google's developer guides, while the open repository focuses on the earlier lines.
A practical route through the code
Four steps that take you from curiosity to a working forecast, in the order that saves the most time.
Decide what you actually need
Direct forecasts, a data feed, or model weights? The first two need no accelerator at all, which changes the whole plan.
Start with the notebook
The quick-start Colab notebook runs the Mini cyclone model end to end on a free TPU runtime: load weights, initialize, roll out a forecast and visualize the outputs.
Or query hosted data
For production use, querying BigQuery or Earth Engine is usually cheaper and simpler than hosting inference. The download page covers the channels.
Then move to an API
When you want forecast fields for a specific location on a schedule, the API path on this site is the shortest route from data to decision.
WeatherNext 3 GitHub questions
Short answers about repositories, weights and access.
Comparing paths? The API page covers request shapes, and the download page covers bulk data.
From repository to hourly forecast
You can run the models from the repository. You can also skip straight to answers: start the 7-day free trial for hourly forecasts anywhere, or read API and data access to wire the data into your own stack.
Related guides
The pages that go with this one, chosen by what each one answers.
