AI Potluck
Back to Gap Map Model components / Scientific AI models

UMA

Meta
open weights / Overall score: 2.4

UMA is Meta FAIR's family of Universal Models for Atoms, trained on roughly half a billion DFT structures compiled from five chemical-domain datasets and routed at inference by a task embedding covering molecules, inorganic materials, catalysis, MOFs, and molecular crystals. Weights are distributed through a gated Hugging Face repository and run through the fairchem library's ASE calculator and multi-GPU simulation interfaces.

Access is gated on manual approval and requires registering a full legal name and organization, which is a conduct and identity condition rather than a commercial bound. Verified 2026-09-10 via the Hugging Face API, the fairchem README, LICENSE and UMA documentation, and the UMA paper.

Openness

4 high confidence
4.0
weights
open(manually gated but downloadable uma-s and uma-m checkpoint files)
data
open(OMol25, OMat24, OC20, ODAC23, and OMC25 training datasets published alongside the model)
code
open(configs/uma/training_release ships uma_sm_direct_pretrain.yaml and uma_md_direct_pretrain.yaml with their backbone, dataset, tasks and element_refs configs, driven by the fairchem CLI, alongside the MIT inference and calculator library)
license
FAIR-Chemistry-License-v1(royalty-free commercial use with an acceptable-use policy and gated registration)

Gating and an acceptable-use policy are identity and conduct conditions rather than a commercial ceiling, so the licence does not push UMA below the open-weights centre of gravity. The pretraining configs in configs/uma/training_release and the five published training datasets are what carry it above the open-weights centre of gravity.

Adoption

1 medium confidence
1.0

83 downloads in the trailing 30 days for the declared Hub repository, with manual gating suppressing the count. fairchem-core is also declared and draws far more, but it is not banded here for a reason rather than by omission: the package is the library that serves every fairchem model, so its download population is not UMA's population, and docs/reference/adoption.md requires a package to be the product's own primary channel before it may band the head product. The Hub repository is UMA's own channel; the gate is what suppresses it.

Capability

3 high confidence
3.0

Two attributes. The five tasks are separate data systems with different functionals and pseudopotentials rather than variables of one, which is what makes this multi-instrument. Adaptation is documented directly - docs/core/common_tasks/fine_tuning.md is a first-party UMA fine-tuning page shipping create_uma_finetune_dataset.py and a templated fine-tuning config, and configs/uma/training_release carries the conserve_finetune YAMLs beside it. The publisher's headline claim is that fine-tuning is usually unnecessary, which is a claim about convenience rather than about whether a pathway exists; the pathway is documented and is counted.

  • https://raw.githubusercontent.com/facebookresearch/fairchem/main/docs/core/uma.md recorded 2026-09-10

    States UMA is trained on five different DFT datasets with different levels of theory, learns a per-task embedding, and requires the user to select a task at inference; tabulates the functional and pseudopotentials for each.

  • https://arxiv.org/abs/2506.23971 recorded 2026-09-10

    Describes training on half a billion 3D atomic structures compiled across multiple chemical domains and reports that a single model without any fine-tuning matches or beats specialized models.

Verified 2026-09-10