MACE

weights

open(ungated foundation checkpoints published as GitHub release assets)

data

open(MPTrj, sAlex, OMat, MATPES, and OMol training datasets publicly released)

code

open(full data preprocessing, distributed training, and fine-tuning pipeline)

license

MIT(code and the MACE-MP-0a, MACE-MP-0b3, and MACE-MPA-0 checkpoints)+ASL(MACE-OMAT-0, MACE-MATPES-PBE-0, MACE-MATPES-r2SCAN-0, MACE-OMOL-0, MACE-OFF23, and MACE-MH-0/1 checkpoints

MACE is otherwise the most complete release in this cohort: open corpus, open pretraining pipeline, open weights. The multi-SKU rule resolves it on the most restrictive licence among the checkpoints actually distributed, and the newer foundation models ship under the Academic Software Licence, which says of itself that it is not an open-source licence because it does not allow commercial use. Publishing the corpus does not buy that permission back, so the score caps at restricted.

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

  Tabulates the recommended foundation models with a download link, a named training dataset, and a licence column per checkpoint — MIT for MACE-MP-0a, 0b3 and MPA-0, ASL for OMAT-0, MATPES and OMOL-0 and MH — and documents mace_run_train, preprocess_data.py, multi-GPU training, and foundation-model finetuning.

• https://raw.githubusercontent.com/ACEsuit/mace/main/LICENSE.md recorded 2026-09-10

  MIT License body with an added ACEsuit/mace copyright line, read directly because the license endpoint reports NOASSERTION; it covers the code, not the checkpoints.

• https://raw.githubusercontent.com/gabor1/ASL/main/ASL.md recorded 2026-09-10

  States the Academic Software Licence is not an open-source licence because it does not allow commercial use, and grants rights only for academic non-commercial use.

Re-banded off stars. mace-torch is a declared artifact, and sources/signal_routing.yaml sets the precedence pypi > huggingface > stars, so an authoritative usage route exists and the record may not fall through to a star count. 113,484 downloads in the trailing 30 days puts it at level 3, one above the 1,345 stars it used to carry. Four first-party Hugging Face repositories also exist under mace-foundations and all report zero downloads, so the Hub route is present but empty; PyPI outranks it either way.

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

  last_month = 113,484 downloads of the mace-torch package.

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

  stargazers_count: 1345; archived: false; fork: false — recorded so the superseded stars signal stays visible.

Two attributes. Most of the family is one checkpoint per domain, and small, medium and large are sizes rather than scales - but MACE-MH is a single distributed checkpoint whose card claims "unified cross-domain capability spanning inorganic crystals, molecular systems, surface chemistry, and reactive organic chemistry with a single model" over heads for OMAT PBE, OMOL, OC20, SPICE, RGD1, MPTraj and MatPES. That is the same fact pattern the category credits UMA for, so it is credited here too.

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

  Documents finetuning a foundation model with mace_run_train --foundation_model, states the pretrained models can be used directly for inference or as a starting point for fine-tuning, and describes outputs as energies, forces and stresses.