UMA, without fine-tuning to specific tasks, performs similarly or better in both accuracy and inference-speed/memory-efficiency than specialized models on a wide-range of material, molecular and catalysis benchmarks.

With all these large diverse datasets, we ask whether it’s possible to train a single model that performs well across several domains without the need for additional finetuning?

Introducing UMA - our latest model trained on over 30 billion atoms across our open-science datasets.

Baseline models are released and evaluated on OOD evaluation tasks. Metrics that go beyond just a structure’s energy and forces are important for understanding how well models perform on downstream applications.

We hope to release a public leaderboard in the near future!

With over 100M structures calculated at ωB97M-V/def2-TZVPD, with variable charge+spin, and spanning biomolecules, electrolytes, metal complexes, and community datasets, OMol25 represents the largest most diverse dataset of its kind.

While our earlier datasets (OC20, OC22, OMat24) focused on catalysts and materials, it was evident the community lacked a large, high quality, diverse molecular DFT dataset for training ML models. The Open Molecules 2025 (OMol25) Dataset hopes to accomplish just that.

Excited to share our latest releases to the FAIR Chemistry’s family of open datasets and models: OMol25 and UMA!
@AIatMeta @OpenCatalyst

OMol25: huggingface.co/facebook/OMo...
UMA: huggingface.co/facebook/UMA
Blog: ai.meta.com/blog/meta-fa...
Demo: huggingface.co/spaces/faceb...