ClimaX

license

MIT(OSI)

weights

open(MIT ungated checkpoints at two grid resolutions)

data

open(Snakemake download and regrid workflow plus preprocessing scripts reconstruct the CMIP6 pretraining corpus end to end)

code

open(pretraining entry point and config alongside per-task fine-tuning configs and data preparation scripts)

ClimaX is the rare complete release in this category - the corpus is rebuildable from shipped scripts rather than merely described, the pretraining run is a documented command with its config, and MIT covers both the repository and the published checkpoints.

• https://microsoft.github.io/ClimaX/usage/ recorded 2026-09-10

  A Snakemake workflow downloads and regrids CMIP6 variables and nc2np_equally_cmip6.py preprocesses them for pretraining; pretraining runs as src/climax/pretrain/train.py --config configs/pretrain_climax.yaml; two pretrained checkpoints at 5.625deg and 1.40625deg are provided, both pretrained on all five CMIP6 datasets, and each downstream task has a fine-tuning config.

• https://raw.githubusercontent.com/microsoft/ClimaX/main/LICENSE recorded 2026-09-10

  Full MIT License text, copyright Microsoft Corporation, with no added use restriction.

• https://huggingface.co/api/models/tungnd/climax recorded 2026-09-10

  license: mit; gated: false; private: false; checkpoint files 5.625deg.ckpt, 1.40625deg.ckpt, vits.ckpt, vitb.ckpt and vitl.ckpt.

The checkpoints sit under tungnd, the personal Hub namespace of Tung Nguyen, ClimaX's first author, and microsoft/ClimaX's own fine-tuning commands point --model.pretrained_path at those exact URLs - so it is the official release rather than a mirror, and it is declared as the weights artifact. It reports zero downloads against 2.6 GB of stored files, so the counter measures the API path rather than the reach, and 709 GitHub stars are the last-resort signal on the aurora-weather precedent.

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

  stargazers_count: 709; archived: false; fork: false.

Two attributes. The named downstream tasks are not counted as an attribute of their own. Multi-instrument is declined on what the inputs are rather than on whether a product source restates the paper - three do, in configs/pretrain_climax.yaml's per-dataset dicts, in arch.py's per-variable tokenization, and in the docs' "pretrained using all 5 CMIP6 datasets". The five are CMIP6 simulation archives, which is one simulation data system under the category's rule that several variables from one data system are not several instruments, and ERA5 enters at fine-tuning rather than pretraining.

• https://microsoft.github.io/ClimaX/usage/ recorded 2026-09-10

  Gives fine-tuning commands and configs for global forecasting, regional forecasting, and climate projection, each loading a pretrained checkpoint, and publishes checkpoints at 5.625deg and 1.40625deg resolution.