Open Molecules 2025 (OMol25)

Open Molecules 2025 (OMol25) is a dataset released by Meta's FAIR Chemistry team, containing over 100 million single-point density functional theory (DFT) calculations computed at the wB97M-V/def2-TZVPD level of theory using ORCA6. Each structure is labelled with total energy (in eV) and atomic forces (in eV per angstrom). The dataset blends elemental, chemical and structural diversity: it covers 83 elements, organic and inorganic molecules, transition metal complexes and electrolytes, explicit solvation, variable charge and spin, conformers and reactive structures, with systems of up to 350 atoms. It represents billions of CPU core-hours of compute.

Access is through HuggingFace, where the dataset is gated behind a request form and provided as ASE DB compatible LMDB files (.aselmdb), with charge and spin multiplicity stored in the atoms.info dictionary. The accompanying fairchem library and documentation cover loading and training workflows, and Meta also released a family of eSEN models trained on the data.

This is a specialist scientific resource rather than a text corpus. It is best suited to training or evaluating machine-learning interatomic potentials and foundation models for chemistry and materials, and for supplying structured facts to a domain RAG system through a metadata layer or knowledge graph rather than raw retrieval. Watch-outs: the dataset licence (CC BY 4.0) requires attribution, access is gated and geographically restricted (unavailable in sanctioned jurisdictions and in China, Russia and Belarus), and the released model checkpoints carry a separate proprietary FAIR Chemistry Licence, so do not assume the models share the data terms. Related sources in this catalogue include Open Materials 2024 (OMat24) for inorganic materials and the broader scientific holdings of the Materials Project.