NeuralGCM

weights

open(CC-BY-SA-4.0 checkpoints on an anonymously readable Google Cloud bucket)

data

documented-not-released(ERA5 inputs, forcings and pressure levels specified with ARCO-ERA5 named as the source

code

partial(inference API, regridding utilities, and reference_code carrying per-checkpoint training configs and loss code, but its entry point imports google_proprietary_code, a package that is not published, so the training path cannot run)

license

CC-BY-SA-4.0(permissive non-OSI)

Apache-2.0 governs the code but the trained weights carry CC-BY-SA-4.0, which is the license that counts here - permissive, share-alike, and not OSI-approved. `code` is partial for a concrete reason rather than the docs' hedge: reference_code ships the per-checkpoint .gin training configs, train_utils.py and the loss code, but experiment.py imports google_proprietary_code, which is not published, so the training path cannot be executed. The docs say the maintainers "would love to support training, modifying and fine-tuning" and are refactoring to make it easier - not that it is unsupported.

• https://raw.githubusercontent.com/neuralgcm/neuralgcm/main/README.md recorded 2026-09-10

  States the split license: code under Apache License 2.0, trained model weights under Creative Commons Attribution-ShareAlike 4.0 International.

• https://neuralgcm.readthedocs.io/en/latest/ recorded 2026-09-10

  "Pseudo-code for training NeuralGCM models can be found in the reference_code subdirectory", the documentation "focuses mostly on our API for inference", and the team notes it would like to support training, modifying and fine-tuning but the present codebase makes it trickier than it needs to be.

• https://neuralgcm.readthedocs.io/en/latest/data_preparation.html recorded 2026-09-10

  Specifies the ERA5 pressure levels, input/output variables and sea-surface forcings the models take, and uses the ARCO-ERA5 Zarr copy of ERA5 as the worked example; no training corpus is distributed.

• https://neuralgcm.readthedocs.io/en/latest/checkpoint_modifications.html recorded 2026-09-10

  Loads checkpoints from gs://neuralgcm/models with an anonymous token, offering v1/deterministic_0_7_deg.pkl, v1/deterministic_1_4_deg.pkl, v1/deterministic_2_8_deg.pkl and v1/stochastic_1_4_deg.pkl.

Re-banded off stars. The record previously said no download count existed; one does. The project's quickstart installs with `pip install -q -U neuralgcm gcsfs`, which makes PyPI provably the product's own channel and its primary one - the two conditions docs/reference/adoption.md sets before a package may band a head product. 8,480 downloads in the trailing 30 days puts it at level 1, one below the 1,014 GitHub stars it used to carry, and adoption.md is explicit that a stars_fallback band re-bands the moment a download signal exists.

• https://pypistats.org/api/packages/neuralgcm/recent recorded 2026-09-10

  last_month = 8,480 downloads of the neuralgcm package.

• https://api.github.com/repos/neuralgcm/neuralgcm recorded 2026-09-10

  stargazers_count: 1014; archived: false; fork: false — recorded so the superseded stars signal stays visible.

Two attributes. Adaptation is explicitly uncounted because the documentation says training and fine-tuning are not supported and ship only as pseudo-code; the checkpoint-modification recipe changes outputs, not learned parameters.

• https://neuralgcm.readthedocs.io/en/latest/checkpoint_modifications.html recorded 2026-09-10

  Distributes deterministic checkpoints at 0.7, 1.4 and 2.8 degrees plus a stochastic 1.4-degree model, and discusses long integrations and the stability of the simulated global mean surface pressure.

• https://neuralgcm.readthedocs.io/en/latest/ recorded 2026-09-10

  Describes NeuralGCM as a hybrid ML and physics atmospheric model for weather and climate simulation built on the Dinosaur dynamical core, and says the maintainers "would love to support training, modifying and fine-tuning NeuralGCM models" but that the Haiku and Gin codebase makes it "much trickier than it needs to be", with pseudo-code for training in reference_code.