Skip to content

Model & framework licenses

oh-my-mlip is an MIT-licensed installer / orchestrator. It does not redistribute any framework's source code — every framework is installed from its own official channel (PyPI / GitHub / Hugging Face / Zenodo) — and every weight is downloaded from its upstream host first. The one exception is a fallback mirror for a few CC-BY-4.0 weights, described below. Each one stays under its own upstream license, which you must comply with — especially for commercial use.

This page is a convenience summary of what we found by reading each project's actual LICENSE file (as of 2026-06). It is not legal advice and licenses change — always check the upstream repo before commercial use or redistribution. Two licenses matter separately:

  • Code license — the framework's Python package (what you pip install).
  • Weights license — the pretrained checkpoint (often different from the code, e.g. a permissive code base with CC-BY or non-commercial weights).

⚠️ Read this first — the restricted ones

Restriction Frameworks (our variants) What it means
🚫 Non-commercial only GRACE (GRACE-2L-OAM), MACE-MH weights (MACE-MH-1-OMAT, MACE-MH-1-OC20) Academic / research use only. Commercial use needs a separate license from the owner.
⚖️ Upstream declarations conflict EquFlash (EquFlashV2, EquFlash-v1), AlphaNet (AlphaNet-v1-OMA) The record hosting the weights says one license and the project says another (details in the table below). Nothing upstream resolves it; ask the authors before relying on either. EquFlash's code is non-commercial (CC BY-NC-SA 4.0); AlphaNet's code is GPL-3.0.
❔ Weights license not stated SevenNet (SevenNet-MF-OMPA, SevenNet-Omni) The checkpoint files oh-my-mlip downloads carry no license statement (details below).
🔗 Weak copyleft (LGPL-3.0) DeePMD (DPA-3.1-3M-FT), DPA4 (DPA-4.0.1-pro-MPtrj) Using the library is fine; modifications to the library must stay LGPL.
🔒 Gated weights UMA (all UMA-* variants), eSEN (eSEN-30M-OAM) Code is MIT, but the weights are under Meta licenses on Hugging Face (FAIR Chemistry License for UMA, OMat24 License for eSEN): accept the license, get access approved, log in with your token. Commercial use is permitted, subject to Meta's Acceptable Use Policy; the FAIR Chemistry License (UMA) also requires compliance with applicable laws, including trade control laws.
📎 Attribution (CC-BY-4.0) weights NequIP (NequIP-OAM-*), Allegro (Allegro-OAM-L), TACE (TACE-OAM-L), DPA weights (DPA-3.1-3M-FT, DPA-4.0.1-pro-MPtrj), MatRIS (MatRIS-10M-OAM, see below) Free incl. commercial, but you must credit the authors.

Everything else below is permissive (MIT / Apache-2.0 / BSD-3-Clause), commercial use allowed.

Full table — 20 frameworks / 32 model variants

Framework Our variant(s) GitHub Code license Weights license Gated
SevenNet SevenNet-MF-OMPA, SevenNet-Omni MDIL-SNU/SevenNet MIT (since 2025-06-20; GPL-3.0 before) Not stated on the GitHub release assets that are downloaded. The authors' figshare deposit of MF-ompa (28590722, a different file) says GPL 3.0+. no
MACE MACE-MPA-0, MACE-MH-1-OMAT, MACE-MH-1-OC20 ACEsuit/mace MIT Mixed: MPA-0 = MIT; MH-1 = ASL (non-commercial) no
NequIP NequIP-OAM-XL, NequIP-OAM-L mir-group/nequip MIT CC-BY-4.0 no
Allegro Allegro-OAM-L mir-group/allegro MIT CC-BY-4.0 no
Nequix Nequix-MP-1 atomicarchitects/nequix MIT MIT (the file is in the MIT-licensed repo); the author's figshare copy of the same file (29927606) says CC BY 4.0 no
DeePMD DPA-3.1-3M-FT deepmodeling/deepmd-kit LGPL-3.0 CC-BY-4.0 (deepmodelingcommunity/DPA-3.1-3M on HF) no
ORB ORB-v3 orbital-materials/orb-models Apache-2.0 Apache-2.0 (code + weights) no
GRACE GRACE-2L-OAM ICAMS/grace-tensorpotential ASL — non-commercial ASL — non-commercial no
MatterSim MatterSim-v1-5M microsoft/mattersim MIT MIT (the 5M checkpoint is downloaded from the GitHub repo on first use) no¹
CHGNet CHGNet-v0.3.0 CederGroupHub/chgnet BSD-3-Clause BSD-3-Clause no
AlphaNet AlphaNet-v1-OMA zmyybc/AlphaNet GPL-3.0 Conflict: the hosting figshare record (28831364) says CC BY 4.0; the author's Matbench Discovery submission says GPL-3.0 no
Eqnorm Eqnorm-MPtrj yzchen08/eqnorm MIT MIT no
fairchem (eSEN) eSEN-30M-OAM facebookresearch/fairchem MIT (code) OMat24 License (Meta; facebook/OMAT24) yes
EquiformerV3 EqV3-OMatMPtrjSalex atomicarchitects/equiformer_v3 MIT MIT (mirror-physics/equiformer_v3 on HF) no
UMA UMA-s-1p2- / UMA-s-1p1- / UMA-m-1p1-* (8) facebookresearch/fairchem MIT (code) FAIR Chemistry License (facebook/UMA) yes
PET PET-OAM-XL metatensor/metatrain BSD-3-Clause BSD-3-Clause (lab-cosmo/upet) no
EquFlash EquFlashV2, EquFlash-v1 SamsungDS/GGNN CC BY-NC-SA 4.0 — non-commercial Conflict: the hosting figshare record (32630910, a Matbench Discovery submission bundle) says CC BY 4.0; the repo license (Samsung Electronics) is CC BY-NC-SA 4.0 and no statement covers the weights no
MatRIS MatRIS-10M-OAM HPC-AI-Team/MatRIS BSD-3-Clause CC BY 4.0 on the hosting figshare record (30475253); the repo and Matbench Discovery declare BSD-3-Clause (both permissive) no
DPA4 DPA-4.0.1-pro-MPtrj deepmodeling/deepmd-kit LGPL-3.0 CC-BY-4.0 (the hosting figshare record 32647224; also declared on Matbench Discovery) no
TACE TACE-OAM-L xvzemin/tace MIT CC-BY-4.0 (xvzemin/tace-foundations) no

¹ MatterSim: only the two open checkpoints (1M / 5M) are MIT; more advanced versions are gated behind Azure Quantum Elements (commercial Microsoft platform). Weights licenses are read from the record that hosts the checkpoint file oh-my-mlip downloads: the Hugging Face license tag, the Zenodo or figshare record's license, or the repository LICENSE when the file lives in the repository. Where another upstream declaration disagrees, the table gives both.

Notes & caveats

  • MACE is the per-model trap. The repo is MIT, but only the Materials-Project line (MACE-MPA-0) ships MIT weights. The MACE-MH-1 weights are under the ASL (Academic Software License) — non-commercial. So of our three MACE variants, only MACE-MPA-0 is commercial-safe.
  • Training-data terms ≠ model terms. CHGNet (MPtrj) and others were trained on datasets with their own terms of use; that constrains re-training/redistributing the data. The OMat24 dataset card licenses the dataset under CC-BY-4.0 and says nothing about models trained on it; each model's own license is what applies.
  • oh-my-mlip never redistributes framework code or gated weights — gated weights (UMA, eSEN) are fetched on first run with your token after you accept the upstream license. See gated_models.md.

Fallback mirror

Files Official source Mirror Why it is allowed
NequIP-OAM-XL-0.1.nequip.zip, NequIP-OAM-L-0.1.nequip.zip, Allegro-OAM-L-0.1.nequip.zip Zenodo 10.5281/zenodo.18775904 ("NequIP & Allegro Foundation Potentials", Kavanagh, S. R.; MIR Group @ Harvard) JinukMoon/oh-my-mlip-mirror-nequip CC-BY-4.0 permits redistribution with attribution; the mirror's model card credits the authors and the record.
eqnorm-mptrj.pt figshare 29153315 ("eqnorm-mptrj ckpt file", Yuzhuo Chen) JinukMoon/oh-my-mlip-mirror-eqnorm The record's license is MIT, which permits redistribution with the copyright notice; the mirror carries the upstream LICENSE text.
MatRIS_10M_OAM.pth.tar figshare 30475253 ("MatRIS-10M checkpoints", Yuanchang Zhou) JinukMoon/oh-my-mlip-mirror-matris The record's license is CC BY 4.0 and the repo's is BSD-3-Clause; both permit redistribution, and the mirror's card credits the authors and carries both notices.
dpa-4.0.1-pro-mptrj.pt figshare 32647224 (DPA-4.0.1-Pro-MPtrj Matbench Discovery submission, Tiancheng Li) JinukMoon/oh-my-mlip-mirror-dpa4 The record's license is CC BY 4.0; the mirror's card credits the author and the record. Only the checkpoint file is mirrored.

The files are byte-identical copies, unchanged. The official host is always tried first (scripts/prepare_nequip_weights.py for Zenodo; scripts/prestage_eqnorm_weights.py, scripts/prestage_matris_weights.py and the registry's weights_mirror_url for figshare), and the mirror is used only when it fails or stays slow. Whichever source delivers a file, it must match the recorded size and checksums, and the log names the source used. Only the files in this table are mirrored.

Quick reference: commercial use

Safe for commercial use (permissive) Restricted — check before commercial use
MACE-MPA-0, NequIP⁴, Allegro⁴, Nequix, ORB, MatterSim (open), CHGNet, Eqnorm, EquiformerV3, PET, MatRIS⁴, TACE⁴, UMA¹, eSEN¹, DeePMD/DPA²,⁴ GRACE (non-commercial), MACE-MH-1 (non-commercial), EquFlash (non-commercial code; weight declarations conflict), AlphaNet (GPL-3.0 code; weight declarations conflict), SevenNet (weights license not stated)

¹ UMA and eSEN: commercial use allowed under Meta's FAIR Chemistry License (UMA) and OMat24 License (eSEN), but the weights are gated and Meta's Acceptable Use Policy applies; the FAIR Chemistry License also requires compliance with applicable laws, including trade control laws. ² DeePMD / DPA4: LGPL-3.0 — commercial use of the library is allowed; copyleft applies only if you modify the library itself. ⁴ CC-BY-4.0 weights (NequIP, Allegro, TACE, DPA, MatRIS) require attribution.

Code licenses read from each upstream LICENSE file (2026-06); weights licenses checked against the record hosting each checkpoint (2026-09-21). Verify upstream before relying on any entry.