WeatherNext
GoogleWeatherNext is Google DeepMind's global medium-range weather and tropical-cyclone forecasting line. One repository now hosts WeatherNext 2, the WeatherNext Cyclones checkpoints and their lightweight Mini variants, together with the earlier GraphCast and GenCast models republished as WeatherNext Graph and WeatherNext Gen, plus a demo notebook that runs forecasts, tracks cyclones, and takes a training step.
The repository absorbed the former google-deepmind/graphcast, so GraphCast and GenCast are versions of this product line rather than separate products.
Openness
3 high confidence- license
- CC-BY-4.0(permissive non-OSI)
- weights
- open(CC-BY-4.0 ungated .npz checkpoints on a public Google Cloud bucket)
- data
- components-listed(ERA5 pretraining and HRES fine-tuning data named and linked through WeatherBench2
- code
- partial(rollout, tracking, loss and a single gradient step in the demo notebook, without a pretraining recipe or configs)
Apache-2.0 covers the notebooks and code, but the weights fall under the README's "all other materials" clause and are therefore CC-BY-4.0, which is permissive without being OSI-approved. The demo goes as far as a gradient step, which is more than inference and short of a released pretraining pipeline.
- https://raw.githubusercontent.com/google-deepmind/weathernext/main/README.md recorded 2026-10-06
Lists downloadable .npz weights for WeatherNext 2, three Cyclones checkpoints and two Cyclones Mini checkpoints on the dm_graphcast Cloud bucket; names ERA5 and WeatherBench2 HRES as the training and fine-tuning data; and states that code is Apache 2.0 while "all other materials are licensed under the Creative Commons Attribution 4.0 International".
Adoption
2 medium confidenceThe weights ship from a Cloud bucket rather than a governed artifact registry, so there is no download count; GitHub stars are the fallback reach signal instead.
- https://api.github.com/repos/google-deepmind/weathernext recorded 2026-09-10
stargazers_count: 7657; archived: false; fork: false.
Capability
4 high confidenceWeatherNext documents three readable capabilities, one fewer than TerraMind documents. Its list of model versions does not count as added breadth, and ERA5 with HRES initial conditions counts as one data system rather than several.
- https://raw.githubusercontent.com/google-deepmind/weathernext/main/README.md recorded 2026-09-10
The demo notebook teaches computing the training loss on predictions and targets and taking a gradient step; WeatherNext 2 is the FGN joint probabilistic architecture and WeatherNext Gen is diffusion-based ensemble forecasting; checkpoints are published at 0.25-degree and, for Mini, 1-degree resolution.
Verified 2026-09-10