COMPANY PROFILE / MATERIALS INFORMATICS
Meta
Large computed datasets for catalysts, materials and molecules, and a "universal" atomic model called UMA, released under conditions
Meta is a social media and advertising company. Its research organisation FAIR (Fundamental AI Research) has a team called "FAIR Chemistry" that, since 2020, has built and released large DFT datasets for catalysts, materials and molecules. In 2025 it announced OMol25, a molecular dataset, and UMA, an atomic AI model trained on these datasets together. We could not find any materials business, product or laboratory facility, and Meta's 10-K does not mention materials. The released items also differ in terms: the data is under CC BY 4.0, while the models carry their own licence, with an access request and prohibited uses. This profile separates the company's claims, the computed data, comparisons with experiment and outside evaluations, and uses only what primary sources let us confirm.
- Meta in 30 seconds, and where it stands in materials informatics
- The logo and how the name is written
- Where the company sits in the materials informatics field
- Scale (the company as a whole, in brief)
- Resources committed to materials informatics
- What it has done so far
- What comes next
- What the company is aiming for in materials informatics
- Map of materials informatics partnerships
- Technology, products and services (how to read the numbers, and outside evaluation)
- The risks this company carries
- Glossary, references and claim-to-source audit
1. Meta in 30 seconds, and where it stands in materials informatics
Materials informatics (MI) means using data, computation and AI to find new materials more efficiently. Its workflow has three broad steps: "data, then prediction and design, then experimental confirmation". Meta mainly works on the data and prediction steps. What sets it apart is that it builds large computed datasets and releases them, and then releases models trained on them.
Nasdaq: METACalifornia, U.S.
OMat24, OMol25, UMADFT data for catalysts, materials and molecules, and models trained on them
OMol25: over 100 millionAs stated in the papers. All are computed data, not experimental values
Models: their own (FAIR Chemistry License)The models require an access request and have prohibited uses
The "right answers" in Meta's MI work are all results of DFT, a physics-based calculation Sourced. Models such as UMA are trained to reproduce DFT answers quickly. They do not learn values measured in experiments directly. So a model's error is usually measured against DFT, and its difference from experiment has to be checked separately. Outside researchers who compared against experimental values have reported sizeable errors. Throughout this profile we keep the company's claims, comparisons with DFT and comparisons with experiment apart.
2. The logo and how the name is written
A note on naming: in this article, "Meta" means Meta Platforms, Inc., the U.S. listed company. The chemistry and materials AI is built by the FAIR Chemistry team within its research organisation FAIR (Fundamental AI Research). Open Catalyst Project is the name of the joint research effort between Meta and Carnegie Mellon University. The 10-K reports figures for the whole company, not for FAIR Chemistry alone.
Official website: https://about.meta.com/ (the logo is the header image of the official site, the white-lettered version made for dark backgrounds)3. Where the company sits in the materials informatics field
As Fig. 1 shows, Meta belongs to the "foundation models and compute" type. Of the MI workflow (data, then prediction and design, then experiment), its centre of gravity is building data through large-scale DFT, releasing it, and releasing models trained on it. The experimental check is not documented in Meta's official materials and is left to users and research partners Sourced.
An AI model that answers questions such as "given this arrangement of atoms, what is the energy and what forces act on each atom?" is called a machine-learning interatomic potential. Traditionally, a separate model was often built for each material or molecule. A universal model aims to work widely across elements with one model. UMA combines different datasets, such as molecules, catalysts, inorganic materials and molecular crystals, in one model, and the user picks a calculation setting (a "task"). For the technical detail, see our explainers on machine-learning interatomic potentials, high-throughput computation, and the basics of density functional theory.
4. Scale (the company as a whole, in brief)
Meta is a large company in social media, advertising, VR and AR. We show only its size as a whole Sourced. Revenue, headcount and budget for materials AI alone are not disclosed Not yet confirmed.
| Metric (whole company, US$ million) | 2023 | 2024 | 2025 |
|---|---|---|---|
| Revenue | 134,902 | 164,501 | 200,966 |
| Research and development | 38,483 | 43,873 | 57,372 |
| Purchases of property and equipment | 27,045 | 37,256 | 69,691 |
| Item | Detail |
|---|---|
| Employees (31 December 2025) | 78,865 |
| Headquarters | 1 Meta Way, Menlo Park, California, U.S. |
| Listing | Nasdaq Global Select Market (META) |
| Statement on open-sourcing AI | The 10-K says the company has a history of releasing models such as Llama, but also that it has not released everything it has developed and expects to keep training a combination of open and closed models |
| Mentions of materials AI | In the 10-K (year to December 2025), we could not find any mention of materials, chemistry, catalysis, FAIR Chemistry or UMA |
The figures above describe the company as a whole. How much of the roughly US$57.4 billion of R&D in 2025 went to materials AI is not disclosed. The 10-K describes large investment in AI computing infrastructure and talent, but does not mention research on materials or chemistry. The size of the company and the size of its commitment to materials AI are two different things.
5. Resources committed to materials informatics
Amounts and headcounts are not disclosed, but we can set out what official sources do and do not show, under four headings.
| Heading | What we could confirm | Type |
|---|---|---|
| People | The UMA paper has 18 authors and the OMat24 paper 9, with the papers presenting the releases as Meta FAIR's. The official documentation names people who contributed to improvements such as UMA Nitro. The total headcount of FAIR Chemistry is not disclosed | Sourced / not disclosed |
| Money | R&D spending and capital expenditure for materials AI are not disclosed. On computing, the official blog says building OMol25 required 6 billion core hours (a company claim) | Could not confirm (only a computing figure is claimed) |
| Investments, acquisitions, joint ventures, divestments | We could not confirm any investment, acquisition, joint venture or divestment concerning materials AI | Could not confirm |
| Universities, research institutes and public funding | (1) Carnegie Mellon University: started the Open Catalyst Project jointly in October 2020. (2) The official blog cites past collaboration with Berkeley Lab, Princeton University, Genentech, Stanford University, the University of Cambridge, CMU, NYU, Los Alamos National Laboratory and UC Berkeley. (3) Lawrence Livermore National Laboratory: collaboration to extend to new molecule classes such as polymers. (4) Data from the COSMO laboratory at EPFL (Switzerland), called MAD, was used to train a fine-tuned UMA variant. We could not confirm receipt of public research funding | Sourced / could not confirm |
Status of the main datasets, models and software
| Name | Kind | Main specifications and conditions | Release and licence | Stage and date |
|---|---|---|---|---|
| OC20 | Dataset for adsorption on catalyst surfaces | 1,281,040 DFT relaxations (about 265 million single-point evaluations). Nitrogen, carbon and oxygen adsorbates (paper) | Open (we did not check the licence for this article) | Released (project started October 2020; paper in 2021) |
| OC22 | Dataset for oxide electrocatalysts | Focused on oxide electrocatalysis (official site) | Open (we did not check the licence for this article) | Released (June 2022) |
| OpenDAC 2023 | Dataset for metal-organic frameworks (MOFs) for direct air capture of CO2 | Centred on CO2 and H2O adsorption (note in the official documentation) | Open (we did not check the licence for this article) | Released (October 2023) |
| OMat24 | DFT dataset of inorganic bulk materials | Over 110 million DFT calculations. Non-equilibrium (rattled) and relaxed structures. PBE/PBE+U with VASP. Training portion of about 100.8 million | Data under CC BY 4.0. The models are under a separate OMat24 License with an access request | Released (October 2024). Validation sets corrected for duplicates in December 2024 |
| OMol25 | DFT dataset of molecules | Over 100 million DFT calculations at ωB97M-V/def2-TZVPD, 83 elements, about 83 million molecular systems, up to 350 atoms. Claimed to need 6 billion core hours | Data under CC BY 4.0. Models under the FAIR Chemistry License v1 | Released (May 2025) |
| UMA | Universal atomic model (machine-learning interatomic potential) | Trained on about half a billion 3D atomic structures. Small (UMA-S: about 6.6 million active and 290 million total parameters) and medium (UMA-M: about 50 million active and 1.4 billion total). A task selects the calculation setting | FAIR Chemistry License v1. Access request on Hugging Face | Released (May to June 2025). UMA-1.2 in March 2026. Nitro and new fine-tuned variants in October 2026 |
| fairchem (library) | Software for using the data and models | A calculator usable from ASE, multi-GPU execution, LAMMPS integration | MIT licence (separate from the licence of the model weights) | Offered (updated continuously) |
There are several points to check before using UMA Sourced.
(1) The "task" decides which DFT is imitated. The omat task for inorganic materials follows PBE/PBE+U (without dispersion correction), and the oc20 task for catalysis follows RPBE. The official documentation notes that omat cannot select a magnetic state and that the training data did not fully cover magnetic states.
(2) The r²SCAN variant is non-magnetic. The fine-tuned variant released in October 2026 (r²SCAN) carries a warning that it treats every system as non-magnetic, and it is advised not to be used for magnetic systems or for cohesive energies.
(3) Terms of use. Access to the models requires a request on Hugging Face, and the licence lists among its prohibited uses those related to military, warfare, the nuclear industry and the U.S. International Traffic in Arms Regulations (ITAR). Published results must acknowledge use of the materials. This is a summary of the official page and not legal advice, so companies should read the original text.
(4) Hardware. The latest speed-up (UMA Nitro) assumes NVIDIA GPUs of the Ampere, Hopper or Blackwell generations.
6. What it has done so far
| When | What happened | What it means |
|---|---|---|
| 14 October 2020 | Meta and Carnegie Mellon University launched the Open Catalyst Project, using AI to look for catalysts for storing renewable energy | The starting point of its MI work |
| 2021 | Published the OC20 dataset paper in ACS Catalysis. 1,281,040 DFT relaxations | Release of a large catalyst dataset |
| 21 June 2022 | Released OC22 (oxide electrocatalysts) | Widened the scope of the data |
| 31 October 2023 | Released OpenDAC (MOFs for direct air capture of CO2) | A new class of materials |
| 16 October 2024 | Announced OMat24 (inorganic bulk materials, over 110 million DFT calculations) and models trained on it, on arXiv | Data for inorganic materials. Claimed a top result on the outside Matbench Discovery benchmark |
| 20 December 2024 | Published corrected OMat24 validation sets because of duplicated structures (a note on the dataset card) | A data correction |
| 14 May 2025 | Announced OMol25 (molecules) and UMA (official blog) | Combined molecules, materials and catalysts into one model |
| 30 June 2025 | Posted the UMA paper on arXiv (trained on about half a billion structures, with a parameter-efficient design) | Release of the method and evaluations |
| October 2025 | Added interfaces for large-scale runs with multiple nodes, multiple GPUs and LAMMPS | Support for large simulations |
| March 2026 | Released UMA-1.2, claimed to be about 50% faster and about 40% more accurate on the Open Molecules test. Extended data to catalysts (oxides and interfaces), molecules and polymers | A company claim. Conditions in section 10 |
| June 2026 | Released the UMA playground, an interactive demo | Easier to try |
| October 2026 | Released UMA Nitro (faster inference), two fine-tuned UMA-S-1.2.1 variants (one for molecules, one for r²SCAN materials) and DFT-D3 dispersion correction | The latest update (GitHub README) |
7. What comes next
The plans we can confirm from official sources centre on extension to polymers and similar classes Not yet confirmed. The company releases its work in a form that suggests it will keep updating it, but it makes no commitment on timing, scale or the name of the next model.
None of the following appeared in official sources Not yet confirmed.
The timing of the next generation of UMA, investment and headcount for materials AI, any laboratory facilities of its own or intake of experimental data, commercial contracts with materials companies, and future changes to the licences.
8. What the company is aiming for in materials informatics
This section includes points the company has not stated directly. We give them as inferences drawn from primary sources, together with the evidence.
| What we read from the sources | Primary evidence | Type |
|---|---|---|
| The aim is not to become a materials company but to build an open research foundation (to become a standard for the research community) | The official blog says the aim is to make the work widely available and foster an open ecosystem. We found no materials product, price or sales statement | Inference |
| It gives data wide freedom but releases models with conditions | Data is under CC BY 4.0; the models carry a licence with an access request and prohibited uses | Sourced (the difference in licences) / Inference (the intent) |
| It aims for one universal model while also issuing task-specific variants | UMA-1.2.1 gained variants for molecules and for r²SCAN materials, which the documentation notes were trained on less data than the original | Inference |
| Within Meta, materials AI is one of FAIR's research themes, not a pillar of the business | The 10-K does not mention it, and its AI coverage centres on Llama, advertising and products | Inference |
The 10-K says that Meta has not released everything it has developed and expects to continue with a combination of open and closed models Sourced. In the materials AI releases too, the data (CC BY 4.0) and the model weights (own licence, access request) are treated differently. From this we read that Meta aims to become a standard for the research community through wide release of data while keeping some control over how the models are used Not yet confirmed.
9. Map of materials informatics partnerships
10. Technology, products and services
What Meta offers for materials is open data (OC20 to OMat24, OMol25), a model (UMA) and software (fairchem) to use them. We could not find any paid service or product. This section explains the path from data to model, the conditions behind the numbers, and outside evaluation.
From data to UMA (illustration)
According to the paper, UMA uses a design called a mixture of linear experts (MoLE). It makes the total number of parameters large while keeping small the number of parameters actually used for each structure (the active parameters). For example, the medium UMA has 1.4 billion parameters in total but about 50 million active per structure Sourced.
How to read the numbers: what was measured, and under what conditions
| Figure (claimed by the company and papers) | Conditions and definition | Nature of the figure |
|---|---|---|
| OMat24: over 110 million DFT calculations | PBE/PBE+U with VASP. It includes many structures away from equilibrium, such as rattled structures and molecular-dynamics snapshots. Training portion of about 100.8 million | Computed data |
| F1 above 0.9 for ground-state stability and 20 meV/atom accuracy for formation energy | The paper claims this for EquiformerV2 models on the outside Matbench Discovery benchmark. The training data was built so that structure types do not overlap with the benchmark's test structures (dataset card) | Comparison with DFT (calculation against calculation) |
| OMol25: over 100 million DFT calculations, 83 elements, up to 350 atoms | Computed at ωB97M-V/def2-TZVPD. About 83 million molecular systems. The official blog says 6 billion core hours | Computed data |
| UMA: trained on about half a billion structures; one model matches or beats specialised models | The paper's claim. Evaluated across several application areas by comparison with DFT. The paper says a single model without fine-tuning performs similarly to or better than specialised models | Comparison with DFT (calculation against calculation) |
| UMA-1.2: about 50% faster, about 40% more accurate on the OMol test | As stated in the GitHub README. For what it is compared against, the original text should be checked | Company claim |
| UMA Nitro: about 6 ns per day (1,000 atoms) | One NVIDIA H200, molecular dynamics through ASE, a 1,000-atom silicon cell. Supports up to 15,000 atoms | A speed measurement (with conditions) |
Most of the accuracy figures above measure agreement with DFT answers. DFT has errors of its own that depend on the functional used (PBE, r²SCAN and so on), and it can disagree with experiment. In UMA the task decides which DFT is imitated, so the same structure can give a different answer under a different task Sourced. Comparison with experimental values needs a separate evaluation.
Outside evaluation and the official cautions (reported neutrally)
| Source | Subject | Main points (as each source states) |
|---|---|---|
| UniFFBench (arXiv, August 2025, revised June 2026) | Universal machine-learning force fields tested on about 1,500 mineral structures with experimental values (including 0 to 5000 K and 0 to 1000 GPa) | For six models (CHGNet, M3GNet, MACE, MatterSim, SevenNet and Orb) it reports a "reality gap": models that score well on computational benchmarks often fail when faced with experimental complexity. In an additional evaluation, UMA and MACE-OMat finish simulations stably, but their density errors (MAPE) of 5.34% and 6.89% (in the order given) exceed practical thresholds, it says. We could not confirm whether it is peer-reviewed |
| UMA's official documentation | Limits on using UMA | States the limits of each task. Examples: omat cannot select a magnetic state, oc20 contains no oxides or explicit solvents, odac centres on CO2 and H2O adsorption, and omol is unlikely to work well for inorganic materials |
| The OMat24 dataset card | Errors in the data | Notes that the validation sets contained duplicated structures and that corrected versions were uploaded on 20 December 2024 |
| Meta's response | The outside evaluation | Within the scope of this check, we could not find an official rebuttal to UniFFBench |
These are matters on which opinions differ, and this article does not decide who is right. What can be confirmed is the following. (1) An outside evaluation reports that sizeable errors remain against experimental values for models including UMA (under conditions that include extreme temperature and pressure). (2) The official documentation itself states the limits of each task. (3) UMA is updated at short intervals, so the version evaluated may differ from the latest.
Three points matter directly to a materials engineer Sourced.
(1) Use it for a fast first look. UMA is orders of magnitude faster than DFT, so it suits narrowing candidates, relaxing structures and previewing molecular dynamics. The practical approach is still to check the important candidates with DFT or experiment before relying on them.
(2) Check that your material system is in range. Read the limits of each task (magnetism, oxides, solvents, surfaces) and check whether your system lies within the training data. Outside that range, a number may come out but cannot be trusted.
(3) Pin and record the version. The models are updated often (UMA-1.1 to 1.2 to 1.2.1). When you report results, naming the model, the version and the task is the basis of reproducibility.
11. The risks this company carries
| Risk | Detail | Type |
|---|---|---|
| The gap with experiment | Training and evaluation are benchmarked against DFT. An outside evaluation reports large errors against experimental values. We could not find an official rebuttal from Meta | Sourced |
| Restrictions on how the models may be used | The models need an access request and have prohibited uses such as military and the nuclear industry. Some companies or fields may not be able to use them | Sourced |
| Continued openness depends on company policy | The 10-K says Meta will not release everything and will keep a mix of open and closed models. We found no commitment to keep releasing for materials | Sourced / Inference |
| Results change with versions | UMA has been updated several times in a little over a year, and library updates have broken compatibility, such as removing the earlier turbo setting (change log). Reproducibility needs care | Sourced |
| Corrections to data | The OMat24 validation sets contained duplicates and were corrected. Results evaluated on the old version need care when compared | Sourced |
| Continued investment in materials AI is invisible from outside | The 10-K does not mention materials AI, and spending and headcount are not disclosed | Inference |
| Limits built into the calculations | The yardstick is DFT with PBE, r²SCAN and similar functionals. Real-world complexity such as temperature, defects and interfaces is limited to what the training data of each task covers | Sourced |
12. Glossary
- Materials informatics (MI)
- Using data, computation and AI to find new materials efficiently.
- FAIR Chemistry
- The team within Meta's research organisation FAIR that builds AI for chemistry and materials.
- Open Catalyst Project
- A joint research effort begun in 2020 by Meta and Carnegie Mellon University to look for catalysts with AI.
- OMat24 / OMol25
- Large datasets of DFT calculations for inorganic materials and for molecules, released by Meta.
- UMA
- Universal Model for Atoms. An atomic AI model that aims to handle molecules, materials and catalysts in one.
- Machine-learning interatomic potential (MLIP)
- A way to calculate the forces between atoms quickly with AI. Used in molecular dynamics.
- DFT (density functional theory)
- A method that calculates the energy of an arrangement of atoms from the behaviour of its electrons. Accurate, but slow.
- Functional (PBE, r²SCAN and others)
- The kind of approximation used inside DFT. The choice changes the result somewhat.
- Task
- The setting in UMA that picks which DFT set-up to imitate (omat, omol, oc20 and so on).
- Matbench Discovery
- An outside benchmark that compares predictions of materials stability.
- CC BY 4.0
- A licence that allows use, including commercial use, with attribution (applied to the data).
- 10-K
- The annual report that U.S.-listed companies file with the SEC.
13. References
- Meta Platforms, Inc. Form 10-K for the fiscal year ended December 31, 2025 (filed with the U.S. SEC on 29 January 2026). Group-wide revenue, R&D, capital expenditure and headcount, and the passage on open-sourcing AI. Also used to confirm that materials, chemistry and catalysis are not mentioned. https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm
- Meta AI Sharing new breakthroughs and artifacts supporting molecular property prediction, language processing, and neuroscience (official blog, 14 May 2025). Release of OMol25 and UMA, collaborators, and the plan to extend to polymers. https://ai.meta.com/blog/meta-fair-science-new-open-source-releases/
- Open Catalyst Project Official site of the Open Catalyst Project. The joint effort between Meta and Carnegie Mellon University, its start in October 2020, and the release dates of OC22 and OpenDAC. https://opencatalystproject.org/
- arXiv / ACS Catalysis Chanussot, L. et al. “The Open Catalyst 2020 (OC20) Dataset and Community Challenges” (arXiv:2010.09990; ACS Catalysis 11, 6059–6072, 2021, DOI 10.1021/acscatal.0c04525). https://arxiv.org/abs/2010.09990
- arXiv Barroso-Luque, L. et al. “Open Materials 2024 (OMat24) Inorganic Materials Dataset and Models” (arXiv:2410.12771, 16 October 2024; revised 20 May 2026). https://arxiv.org/abs/2410.12771
- Hugging Face Dataset card of facebook/OMAT24. Contents, splits, licence (CC BY 4.0), and the note on the corrected validation sets (20 December 2024). https://huggingface.co/datasets/facebook/OMAT24
- arXiv Levine, D. S. et al. “The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models” (arXiv:2505.08762, 13 May 2025; revised 4 March 2026). https://arxiv.org/abs/2505.08762
- Hugging Face Page of facebook/OMol25. Data under CC BY 4.0, models under the FAIR Chemistry License v1, and the prohibited uses in the acceptable use policy. https://huggingface.co/facebook/OMol25
- arXiv Wood, B. M. et al. “UMA: A Family of Universal Models for Atoms” (arXiv:2506.23971, 30 June 2025; revised 4 March 2026). https://arxiv.org/abs/2506.23971
- Hugging Face Model page of facebook/UMA. Gated access, the FAIR Chemistry License v1 (updated 14 May 2025) and the acceptable use policy. https://huggingface.co/facebook/UMA
- Meta FAIR Chemistry (GitHub) README of facebookresearch/fairchem. Licence of the library (MIT), latest news (June 2025 to October 2026), list of models and description of tasks. https://github.com/facebookresearch/fairchem
- FAIR Chemistry Documentation UMA Models (official documentation). The UMA tasks with the level of theory and usage notes for each, and the limitations of the UMA-S-1.2.1 fine-tuned variants. https://facebookresearch.github.io/fairchem/uma
- arXiv Mannan, S. et al. “UniFFBench: Evaluating Universal Machine Learning Force Fields Against Experimental Measurements” (arXiv:2508.05762, 7 August 2025; revised June 2026; we could not confirm whether it is peer-reviewed). An outside evaluation. https://arxiv.org/abs/2508.05762
- Meta Meta's official website (source of the logo: the header image on the official site). https://about.meta.com/
14. Claim-to-source audit
| Claim in the text | Type | Source |
|---|---|---|
| Meta's revenue, R&D and purchases of property and equipment for 2023 to 2025, 78,865 employees, headquarters and listing | Sourced | Reference 1 https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm |
| The 10-K does not mention materials, chemistry, catalysis, FAIR Chemistry or UMA; its statement that it has not released everything it has developed | Sourced | Reference 1 https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm |
| Revenue, headcount and budget for materials AI are not disclosed | Not yet confirmed | Not in the 10-K. Reference 1 https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm |
| Announcement of OMol25 and UMA (14 May 2025), collaborators, 6 billion core hours, extension with LLNL, aim of an open ecosystem | Sourced | Reference 2 https://ai.meta.com/blog/meta-fair-science-new-open-source-releases/ |
| The Open Catalyst Project is a joint effort of Meta and Carnegie Mellon University (started 14 October 2020); OC22 (June 2022); OpenDAC (October 2023) | Sourced | Reference 3 https://opencatalystproject.org/ |
| OC20: 1,281,040 DFT relaxations and about 265 million single-point evaluations | Sourced | Reference 4 https://arxiv.org/abs/2010.09990 |
| OMat24: over 110 million DFT calculations, EquiformerV2, F1 above 0.9 and 20 meV/atom, claimed result on Matbench Discovery | Sourced | Reference 5 https://arxiv.org/abs/2410.12771 |
| OMat24 dataset: about 100.8 million training entries, CC BY 4.0, no overlap with the benchmark's test structures, correction of validation sets (20 December 2024) | Sourced | Reference 6 https://huggingface.co/datasets/facebook/OMAT24 |
| OMol25: over 100 million DFT calculations, omega B97M-V/def2-TZVPD, 83 elements, about 83 million systems, up to 350 atoms | Sourced | Reference 7 https://arxiv.org/abs/2505.08762 |
| OMol25 data under CC BY 4.0, models under the FAIR Chemistry License v1, prohibited uses (military, the nuclear industry, ITAR and others), access request | Sourced | Reference 8 https://huggingface.co/facebook/OMol25 |
| The UMA paper: about half a billion structures, a single model matching or beating specialised models (claimed), medium model with 1.4B total and about 50M active parameters, MoLE design | Sourced | Reference 9 https://arxiv.org/abs/2506.23971 |
| The UMA page on Hugging Face: access request, FAIR Chemistry License v1 (updated 14 May 2025), acknowledgement of use and compliance with trade control laws and the policy | Sourced | Reference 10 https://huggingface.co/facebook/UMA |
| MIT licence of fairchem, UMA-1.2 (March 2026), playground (June 2026), UMA Nitro, new variants and D3 correction (October 2026), model list and parameter counts, speed conditions, multi-GPU and LAMMPS (October 2025) | Sourced | Reference 11 https://github.com/facebookresearch/fairchem |
| UMA tasks with level of theory and usage notes, non-magnetic r2SCAN variant, training on EPFL's MAD data, variants trained on less data | Sourced | Reference 12 https://facebookresearch.github.io/fairchem/uma |
| UniFFBench: about 1,500 minerals, six models, "reality gap", UMA and MACE-OMat density errors of 5.34% and 6.89% | Sourced | Reference 13 https://arxiv.org/abs/2508.05762 |
| The logo image (header image of the official site, for dark backgrounds) | Sourced | Reference 14 https://about.meta.com/ |
| Timing of the next UMA, materials AI spending and headcount, own laboratory facilities, commercial contracts, future licence changes | Not yet confirmed | Not in official sources. Not estimated in this article. Reference 11 https://github.com/facebookresearch/fairchem |
| Investments, acquisitions, joint ventures or divestments in materials AI, and receipt of public research funding | Not yet confirmed | Could not confirm. We do not claim they do not exist either. Reference 1 https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm |
| The foundation-models-and-compute type and the arrangement of Fig. 1 and Fig. 2 | Our calculation | Our own arrangement. Company names are examples from the companies covered in this series |
| The reading that the aim is to build an open research foundation | Inference | Our own interpretation of the official blog and of the absence of products and prices |
| The reading that data is released freely but models with conditions | Inference | Our own interpretation of the difference in licences |
| The reading that it aims for one universal model while also issuing task-specific variants | Inference | Our own interpretation of the variants added in UMA-1.2.1 |
| The reading that materials AI is one of FAIR's research themes and not a pillar of the business | Inference | Our own interpretation of the absence of materials AI in the 10-K |
Last updated 22 October 2026 / Troy Technical
The figures in this article come from Meta's 10-K filed with the U.S. Securities and Exchange Commission, Meta AI's official blog, FAIR Chemistry's GitHub and documentation, papers on arXiv and elsewhere, and the licence notices on Hugging Face. We do not use market estimates from research firms or figures from news reports. Passages marked "Inference" are our own readings of primary sources and are not statements by the company. Accuracy figures are presented as claims by the company or research group, and we keep comparison with DFT apart from comparison with experiment. The descriptions of licences summarise the official pages and are not legal advice. We say nothing about share prices, valuation or investment decisions.