COMPANY PROFILE / MATERIALS INFORMATICS
IBM
— Publishing materials AI as open foundation models, and putting them to work with partners
IBM as a whole is a hybrid-cloud and enterprise-AI company, and in the Form 10-K and annual report for 2025 we found no mention of materials discovery. We cover it here because its research division, IBM Research, has published several chemistry and materials foundation models under the Apache-2.0 license on GitHub and Hugging Face. A preprint also reports using one of them to search for battery electrolytes: 16 formulations were actually measured, and 7 passed the study’s threshold.
This profile ignores IBM as a whole and focuses on its materials AI research: what it has released, who it works with, and what we could not confirm. Terms such as “foundation model” and “molecular representation” are explained where they first appear.
- IBM 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 whole company 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
- The risks this company carries
- Glossary
- References
- Claim-to-source audit
1. IBM in 30 seconds, and where it stands in materials informatics
(primary)Secondary: autonomous experiments (RoboRXN). Current status not confirmed
for chemistry and materialsFM4M (Foundation Models for Materials)
Materials informatics (MI) is a way of finding new materials by collecting materials data, using AI or computation to predict and design properties, and then checking the result by experiment. IBM sits on the “predict and design” step: it builds AI models for it and releases them free of charge Sourced. It is not selling them as a product. The public record shows a research division that publishes models and then puts them to work with materials companies such as JSR (our reading is in Chapter 8).
But IBM’s overall scale and the scale of its materials AI research are two different things. IBM does not disclose headcount, budget or revenue for materials AI research Not yet confirmed.
2. The logo and how the name is written
A note on names: the legal entity is International Business Machines Corporation, abbreviated IBM. IBM Research, which does the materials AI work, is not a separate company but IBM’s research division. This article writes “IBM” for the whole company and “IBM Research” for the division.
A note on model names: IBM’s blog writes SMILES-TED, while the model card and the paper write SMI-TED. They appear to be the same model, and this article follows the model card and paper (SMI-TED).
Official site: https://www.ibm.com/3. Where the company sits in the materials informatics field
Even when two companies both say “AI”, they differ in where they sit along the flow of collect data → predict and design → check by experiment. Here is IBM placed on that flow.
For the basics of MI, see our explainer “Introduction to Materials Informatics”; for how molecules are fed to machines, see “Materials Data and Descriptors” and “Graph Neural Networks”.
4. Scale (the whole company in brief)
Only IBM’s overall size is given here Sourced.
| Item | FY2024 (to Dec 2024) | FY2025 (to Dec 2025) |
|---|---|---|
| Revenue (US$ million) | 62,753 | 67,535 |
| R&D expense (US$ million) | 7,479 | 8,316 |
| Item | Detail |
|---|---|
| Employees (31 December 2025) | IBM and wholly owned subsidiaries: 264.3 thousand (8.7 thousand in less-than-wholly owned subsidiaries and 13.8 thousand complementary workforce are shown separately) |
| Headquarters | Armonk, New York, USA (an owned site) |
| Listing | New York Stock Exchange (NYSE) |
These figures are for the whole company. They do not show the scale of materials AI research. The US$8,316 million of R&D is company-wide, and the annual report attributes the increase mainly to investment in AI, hybrid cloud and quantum, with no materials breakdown. IBM discloses none of the revenue, investment or dedicated headcount for materials AI research Not yet confirmed.
5. Resources committed to materials informatics
Amounts and headcounts are not disclosed, but the official sources do show what is being put in. Items we could not confirm are marked as not found.
| Type of resource | What we found | Type |
|---|---|---|
| People | Dedicated headcount for materials AI research is not disclosed. IBM’s blog names Seiji Takeda (principal scientist) and Emilio Vital Brazil (senior scientist) as co-leads of FM4M. Official sources point to research sites in Tokyo (the director of IBM Research – Tokyo appears in JSR’s release) and Zurich (the 2021 RXN release, and a 2022 blog post by IBM Research Europe) | Sourced (headcount not disclosed) |
| Money | Company-wide R&D expense was US$8,316 million in 2025. Budget, capital spending and compute investment for materials AI research are not disclosed | Sourced (no breakdown) |
| Equity stakes, acquisitions, joint ventures, divestitures | The Form 10-K and annual report contain no mention of equity stakes, acquisitions, joint ventures or sales related to materials AI | Not found |
| Universities, research institutes and public funding | The AI Alliance working group for materials and chemistry (WG4M) has about 20 corporate and academic participants (IBM blog, December 2024). An agreement with JSR on a joint AI research program (November 2025). Public funding such as DOE, NSF or NEDO is not mentioned in the sources we checked | Sourced (public funding not found) |
Status of the main models and products
These are the main materials and chemistry models that IBM Research has published. Dates are repository creation dates on Hugging Face or GitHub (not necessarily the day public use began) or paper dates. Performance figures are claims by the company or study team and were not reproduced for this article.
| Name | Kind | Main specifications (published training data, target) | Release form and license | Stage | Date | Source |
|---|---|---|---|---|---|---|
| SMI-TED | Chemical language model (encoder-decoder) | Trained on about 91 million SMILES from PubChem (about 4 billion tokens). Two variants, 289M and 8x289M. The paper reports that it matches or exceeds existing approaches on benchmarks | Apache-2.0; weights public | Released | HF 25 Jul 2024; paper 1 Jul 2025 | Reference 5, 14 |
| SELFIES-TED | Chemical language model (SELFIES) | IBM’s blog says about 1 billion SELFIES samples (see the note below on data sources) | Apache-2.0 | Released | HF 25 Oct 2024 | Reference 7, 3 |
| MHG-GED | Molecular-graph model | Trained on about 1.34 million molecules from PubChem (the blog says about 1.4 million) | Apache-2.0 | Released | HF 25 Oct 2024 | Reference 6, 3 |
| MoL-MoE | Mixture of experts that combines three representations | 12 experts: four each for SMILES, SELFIES and molecular graphs. IBM reports it beat single-representation models on MoleculeNet classification and regression (split and metric details are not in the blog) | Apache-2.0 | Released | HF 12 Jun 2025 | Reference 8, 3 |
| pos-egnn | Graph model that handles atomic positions (foundation model) | Trained on about 1.4 million samples (90%) of the Materials Project Trajectory (MPtrj) dataset to predict energies, forces and stress | Apache-2.0 | Released | HF 21 Mar 2025 | Reference 9 |
| Geodite | Architecture for a universal machine-learning interatomic potential | Avoids tensor products to cut computation. The card shows how to use Geodite-MP, trained on MPtrj. The citation field on the card reads “TODO”, and we could not find a paper | Apache-2.0 | Released (paper not found) | HF 3 Nov 2025 | Reference 10 |
| STR-Bamba | Encoder-decoder for many text forms of molecules | Trained on 588 million samples (about 29 billion tokens). Handles SMILES, SELFIES, molecular formula, InChI, IUPAC name, polymer SMILES and formulations | Apache-2.0 | Released | HF 11 Dec 2025 | Reference 11 |
| GT4SD | Toolkit for generative models | An open-source toolbox for generating candidate molecules and materials | MIT | Released | GitHub created 11 Feb 2022 | Reference 13, 4 |
| RXN for Chemistry / RoboRXN | AI for reaction prediction, and a cloud lab that synthesizes with robots | The 2021 release says it began in 2018, with more than 29,000 users, over 5 million reaction predictions and over 90% accuracy on forward predictions (company claims). We could not confirm the 2026 operating status | Delivery form not confirmed | As stated in 2021 | Nov 2021 | Reference 18 |
IBM’s blog says SMILES-TED and SELFIES-TED were trained on data “from PubChem and Zinc-22”. The SMI-TED paper and model card, however, say 91 million samples from PubChem and do not mention Zinc-22. Because details may differ, this article gives priority to the primary source for each model (paper or model card) and marks blog statements as “according to the blog”.
6. What it has done so far
| When | What happened | Why it matters |
|---|---|---|
| 2018 | RXN for Chemistry (AI reaction prediction) began (per the 2021 release) | The work started in chemical synthesis rather than in materials as such |
| Nov 2021 | IBM announced RoboRXN links with Atinary, Arctoris, Chemspeed Technologies, Syngenta and Thieme Chemistry | Connections to synthesis automation, reaction data and lab equipment. Current status not confirmed |
| Jun 2022 | An IBM blog post described AI inverse design of CO2-separation-membrane monomers and high-throughput computation for battery electrolytes | The blog itself separates the computed estimates from molecular dynamics from experimental confirmation |
| Spring 2024 | IBM and JSR launched the AI Alliance working group for materials and chemistry (WG4M), per IBM’s blog | About 20 corporate and academic participants working on foundation models, datasets and benchmarks |
| Jul 2024 | FM4M release began (GitHub repository created 24 July; SMI-TED registered on Hugging Face 25 July) | The start of open release of materials models |
| Oct 2024 | SELFIES-TED and MHG-GED released | The three molecular representations were now all covered |
| 20 Dec 2024 | IBM’s blog reported more than 100,000 downloads in a few months | A company statement. Downloads do not directly show use or results |
| Mar 2025 | A preprint on electrolyte search appeared (pos-egnn was also registered) | An example of a public model applied to a real materials search |
| 1 Jul 2025 | The SMI-TED paper was published in Communications Chemistry | Performance reported in a peer-reviewed paper |
| 30 Aug 2025 | The electrolyte paper (IBM’s publication page lists npj Computational Materials) | We did not check the journal version |
| Nov 2025 | JSR and IBM agreed on a joint AI research program (4 November). Geodite (3 November) and STR-Bamba (11 December) were registered | The joint research is at the announcement stage; no results are stated |
7. What comes next
JSR’s release describes working with IBM on an “expanded foundation model for materials” and a data platform built with generative AI, to create a platform that handles several materials systems at once Not yet confirmed. It states no funding, period or success measure.
IBM’s blog (December 2024) says it plans, “in the coming year”, to release new fusion techniques and models built on additional data modalities, including the positions of atoms in 3D space Not yet confirmed. pos-egnn, which handles atomic positions, was registered on Hugging Face in March 2025 Sourced.
None of the following is stated in the official sources we checked Not yet confirmed.
The budget and headcount for materials AI research; the next release date for materials models; any paid service for materials AI; the term and deliverables of the JSR research; and RoboRXN’s current availability. Press-based talk of IBM launching a materials product in a given year is not used in this article.
8. What the company is aiming for in materials informatics
This part includes things the company does not say directly. It is our reading of the primary sources, written with the evidence for each point.
| What we read from it | Primary evidence | Type |
|---|---|---|
| Materials AI is mainly open publication of research results, not a product | The models are public on GitHub and Hugging Face under Apache-2.0. The Form 10-K and annual report have no materials-discovery section, and we found no paid offering | Inference |
| It works by putting the models to use with partner companies | The AI joint research program with JSR, and the AI Alliance WG4M (about 20 companies and universities) | Inference |
| Fine-tuning on small data is the practical focus | The blog says anyone with a small amount of data can customize the models. In the electrolyte example, 13,666 measurements were used for fine-tuning | Inference |
| Combining several molecular representations is a pillar of the technology | Three models (SMI-TED, SELFIES-TED, MHG-GED), MoL-MoE, and STR-Bamba’s multi-representation support | Sourced |
| It is extending to machine-learning interatomic potentials that handle atomic positions directly | pos-egnn (March 2025) and Geodite (November 2025) were released. We found no paper for Geodite | Inference |
In the electrolyte search, the public model (SMI-TED) was fine-tuned on 13,666 literature values, about 100,000 candidates were screened by computation, and 16 formulations were measured Sourced. The setup shows a public model’s value as something a company can tailor with a small amount of its own data, not only something for those with huge datasets. From this we read IBM as aiming to provide a base on which materials companies grow models from their own data Not yet confirmed.
9. Map of materials informatics partnerships
10. Technology, products and services
Three ways to feed molecules to AI
To train AI on molecules, you first need to turn each molecule into a form a computer can read. IBM builds a separate model for each way of writing a molecule and then mixes them (according to IBM’s blog).
| Representation | What it is | Strength | Weakness (per IBM’s blog) |
|---|---|---|---|
| SMILES | A molecule written as one line of text (ethanol is CCO) | Being text, it can use language-model technology directly. Huge databases exist | It loses 3D information, so AI can produce invalid molecules |
| SELFIES | A refined text form whose grammar always gives a valid molecule | Broken molecules are less likely | It also carries no 3D information |
| Molecular graph (MHG) | Atoms as points and bonds as lines | Handles how atoms connect directly | Heavy computation |
The three models (SMI-TED, SELFIES-TED, MHG-GED) are each pre-trained on their own representation. MoL-MoE uses a “mixture of experts”: for each query it picks the experts best suited and blends their answers. According to IBM’s blog, the model that mixes the three representations beat single-representation models on MoleculeNet (a benchmark set from Stanford for predicting molecular properties), on classification (for example whether a molecule is toxic) and regression (for example how well it dissolves in water). However, the blog gives no data split, metric or numbers, so this article treats the comparison as the company’s claim and does not rank models.
An example: searching for electrolytes with a public model
IBM’s researchers applied the public model (SMI-TED) to battery electrolytes (as stated in preprint v2; these are the study team’s claims).
The reported numbers are as follows Sourced (all are the study team’s reports).
| Item | What the preprint states | How to read it |
|---|---|---|
| Fine-tuning data | 13,666 ionic-conductivity values collected from the lithium-ion battery literature (non-aqueous electrolytes) | A patchwork of data measured under different conditions |
| Number of candidates | About 100,000 computer-generated formulations predicted with the fine-tuned model | This is a computed prediction, not an experiment |
| Number measured | 16 formulations chosen and measured for ionic conductivity at room temperature | Chosen for high predicted conductivity and salt diversity |
| Number above the threshold | 7 of 16 above 10 mS/cm (a 44% success rate) | The threshold is the conductivity of 1 M LiPF6 in EC/DMC/EMC |
| Best LiFSI case | 19.14 mS/cm, 82% above the best in the literature set. The model predicted 10.52 mS/cm, an underprediction | “Best” is relative to the same literature dataset |
| LiDFOB cases | A formulation 172% more conductive than the best in the literature | We did not check the individual value |
Three points matter to a materials engineer here Sourced. (1) The gap between prediction and measurement: for the best formulation, the prediction was 10.52 and the measurement 19.14 mS/cm, a large mismatch. The paper says the trend was validated, which also means the numbers themselves cannot be taken at face value. (2) The narrow set of properties: only ionic conductivity was measured; cell life, safety, compatibility with electrodes and cost are separate questions. A highly conductive formulation is not automatically a usable electrolyte. (3) Data quality: the 13,666 values are a patchwork from the literature. If you try the same approach with your own data, more data measured under uniform conditions may shrink the gap between prediction and measurement (an inference).
The paper itself also says that correlations among variables were mostly weak and uninformative, and that interpretation methods are needed. Whether a person can make chemical sense of why the AI picked a formulation remains an open problem.
11. The risks this company carries
| Risk | Detail | Type |
|---|---|---|
| Materials effort and investment are invisible | Headcount, budget and revenue are all undisclosed, so progress is hard to check from outside | Sourced |
| Comparison conditions are limited | The MoL-MoE comparison has no numbers in IBM’s blog, and the data split and metric cannot be checked | Sourced |
| Experimental confirmation is limited | The measured results are mainly ionic conductivity at room temperature for 16 formulations; we found no independent reproduction | Sourced |
| Some models lack a paper | Geodite’s model card has “TODO” in the citation field, and we found no paper | Sourced |
| RoboRXN’s status is unclear | We could not confirm operations after the 2021 figures (users, predictions) | Sourced |
| Joint research is at the plan stage | The JSR research is at the announcement stage with no term or deliverables stated | Not yet confirmed |
| Sources disagree | The blog and the model cards or papers describe the origin of training data differently | Sourced |
| No visible path to a paid service | Open release is the focus, and no monetization policy is confirmed | Inference |
12. Glossary
- Materials informatics (MI)
- Finding new materials with materials data plus AI and computation, in the flow of data, then predict and design, then experiment.
- Foundation model
- A large AI model pre-trained on a great deal of data, which can then be adapted to another use with a small amount of data.
- Fine-tuning
- Training a pre-trained model further on your own data to suit your use.
- SMILES / SELFIES
- Ways to write a molecule as text. SELFIES is a refinement that is less likely to produce broken molecules.
- Molecular graph
- A representation of a molecule with atoms as points and bonds as lines.
- Encoder-decoder
- A model made of a part that turns input into a list of numbers (an embedding) and a part that turns it back.
- Mixture of experts (MoE)
- A scheme that picks, case by case, the most suitable among several “expert” models.
- Machine-learning interatomic potential (MLIP)
- A machine-learning model that predicts energy and forces from atomic positions faster than quantum calculations.
- Ionic conductivity
- How easily an electrolyte lets ions pass. The unit is mS/cm.
- Inverse design
- Starting from the properties you want and searching for molecules or materials that have them.
- MoleculeNet
- A public benchmark collection for comparing models that predict molecular properties.
- Form 10-K
- The annual report that US-listed companies file with the SEC.
13. References
- IBM Form 10-K for the year ended 31 December 2025 (filed 24 February 2026). Headquarters, listing and business description; checked for any mention of materials discovery. https://www.sec.gov/Archives/edgar/data/51143/000005114326000010/ibm-20251231.htm
- IBM 2025 Annual Report to Stockholders (Exhibit 13 to the Form 10-K above). Revenue, R&D expense and headcount. https://www.sec.gov/Archives/edgar/data/51143/000005114326000010/ibm-20251231_d2.htm
- IBM Research Blog post “Meet IBM’s new family of AI models for materials discovery” (20 December 2024). FM4M overview, training-data sizes, mixture of experts (MoE), and the AI Alliance WG4M. https://research.ibm.com/blog/foundation-models-for-materials
- IBM Research Blog post “new sustainable materials” (27 June 2022). AI for CO2 membranes, battery electrolytes and biodegradable plastics, with computed and experimental results kept apart. https://research.ibm.com/blog/new-sustainable-materials
- IBM Research Hugging Face model card materials.smi-ted (SMI-TED). Training data, model sizes and license (Apache-2.0). https://huggingface.co/ibm-research/materials.smi-ted
- IBM Research Hugging Face model card materials.mhg-ged (MHG-GED). Training data and license. https://huggingface.co/ibm-research/materials.mhg-ged
- IBM Research Hugging Face model card materials.selfies-ted (SELFIES-TED). https://huggingface.co/ibm-research/materials.selfies-ted
- IBM Research Hugging Face model card materials.mol_moe (MoL-MoE). The 12-expert layout. https://huggingface.co/ibm-research/materials.mol_moe
- IBM Research Hugging Face model card materials.pos-egnn (pos-egnn). Training data (MPtrj) and license. https://huggingface.co/ibm-research/materials.pos-egnn
- IBM Research Hugging Face model card materials.geodite (Geodite). Architecture of a machine-learning interatomic potential. https://huggingface.co/ibm-research/materials.geodite
- IBM Research Hugging Face model card materials.str-bamba (STR-Bamba). Training-data size and supported molecular representations. https://huggingface.co/ibm-research/materials.str-bamba
- IBM GitHub IBM/materials (FM4M). License (Apache-2.0) and repository record. https://github.com/IBM/materials
- GT4SD (originating from IBM Research) GitHub GT4SD/gt4sd-core. Generative-model toolkit. License (MIT). https://github.com/GT4SD/gt4sd-core
- Soares et al. “An open-source family of large encoder-decoder foundation models for chemistry”, Communications Chemistry 8, 193 (1 July 2025, DOI 10.1038/s42004-025-01585-0). https://www.nature.com/articles/s42004-025-01585-0
- Zohair et al. “Chemical Foundation Model Guided Design of High Ionic Conductivity Electrolyte Formulations”, arXiv:2503.14878 (March 2025, v2). The electrolyte search procedure and measured results. https://arxiv.org/abs/2503.14878
- IBM Research IBM publication page for the electrolyte paper above. It lists the venue as npj Computational Materials (30 August 2025). https://research.ibm.com/publications/chemical-foundation-model-guided-design-of-high-ionic-conductivity-electrolyte-formulations
- JSR News release (4 November 2025). JSR and IBM agree to start a joint AI research program. https://jsr.co.jp/jsr_e/news/2025/20251104.html
- IBM Newsroom release “IBM Expands AI-Driven Materials Discovery Capabilities and Signs New Industry Collaborations” (29 November 2021). RXN for Chemistry, RoboRXN and partners. https://uk.newsroom.ibm.com/2021-11-29-IBM-Expands-AI-Driven-Materials-Discovery-Capabilities-and-Signs-New-Industry-Collaborations
- IBM Official website (source of the logo; used to check the company profile). https://www.ibm.com/
14. Claim-to-source audit
| Claim in the text | Type | Source |
|---|---|---|
| Legal name, headquarters and listing | Sourced | Reference 1 https://www.sec.gov/Archives/edgar/data/51143/000005114326000010/ibm-20251231.htm |
| Revenue and R&D expense (2024 and 2025); 264.3 thousand employees | Sourced | Reference 2 https://www.sec.gov/Archives/edgar/data/51143/000005114326000010/ibm-20251231_d2.htm |
| The Form 10-K and annual report contain no materials-discovery mention, as far as we checked | Sourced | Reference 1 https://www.sec.gov/Archives/edgar/data/51143/000005114326000010/ibm-20251231.htm Reference 2 https://www.sec.gov/Archives/edgar/data/51143/000005114326000010/ibm-20251231_d2.htm |
| FM4M model names, training-data sizes and release form (Apache-2.0) | Sourced | Reference 3 https://research.ibm.com/blog/foundation-models-for-materials Reference 5 https://huggingface.co/ibm-research/materials.smi-ted Reference 6 https://huggingface.co/ibm-research/materials.mhg-ged Reference 7 https://huggingface.co/ibm-research/materials.selfies-ted Reference 8 https://huggingface.co/ibm-research/materials.mol_moe Reference 9 https://huggingface.co/ibm-research/materials.pos-egnn Reference 10 https://huggingface.co/ibm-research/materials.geodite Reference 11 https://huggingface.co/ibm-research/materials.str-bamba Reference 12 https://github.com/IBM/materials |
| SMI-TED trained on 91 million SMILES (from PubChem) and published in Communications Chemistry | Sourced | Reference 14 https://www.nature.com/articles/s42004-025-01585-0 Reference 5 https://huggingface.co/ibm-research/materials.smi-ted |
| Three molecular representations, the MoE mixing, and the MoleculeNet comparison (company claim) | Sourced | Reference 3 https://research.ibm.com/blog/foundation-models-for-materials |
| The blog and the model cards or papers differ on where the training data came from | Sourced | Reference 3 https://research.ibm.com/blog/foundation-models-for-materials Reference 5 https://huggingface.co/ibm-research/materials.smi-ted Reference 14 https://www.nature.com/articles/s42004-025-01585-0 |
| Electrolyte search: 13,666 values, about 100,000 predictions, 16 measured, 7 above threshold, LiFSI case 19.14 mS/cm (the study team’s report) | Sourced | Reference 15 https://arxiv.org/abs/2503.14878 |
| IBM’s publication page lists the electrolyte paper under npj Computational Materials (30 August 2025) | Sourced | Reference 16 https://research.ibm.com/publications/chemical-foundation-model-guided-design-of-high-ionic-conductivity-electrolyte-formulations |
| JSR and IBM agreed on a joint AI research program (4 November 2025) and its aims | Sourced | Reference 17 https://jsr.co.jp/jsr_e/news/2025/20251104.html |
| AI Alliance WG4M (launched by IBM and JSR in spring 2024, about 20 participants) | Sourced | Reference 3 https://research.ibm.com/blog/foundation-models-for-materials |
| The 2021 RXN and RoboRXN announcement, partners (Atinary, Arctoris, Chemspeed, Syngenta, Thieme) and usage figures | Sourced | Reference 18 https://uk.newsroom.ibm.com/2021-11-29-IBM-Expands-AI-Driven-Materials-Discovery-Capabilities-and-Signs-New-Industry-Collaborations |
| The 2022 blog: monomer performance is a computed molecular-dynamics estimate | Sourced | Reference 4 https://research.ibm.com/blog/new-sustainable-materials |
| GT4SD license (MIT) and repository creation date | Sourced | Reference 13 https://github.com/GT4SD/gt4sd-core |
| Placement on the three steps of MI (data, predict and design, experiment) | Our calculation | Our own grouping from public sources (this article) |
| Revenue, investment and dedicated headcount for materials AI research | Not yet confirmed | None are disclosed, and this article does not estimate them (not stated in the sources checked for this article) |
| Equity stakes, acquisitions and public funding (DOE, NSF, NEDO and so on) for materials AI | Not yet confirmed | Not stated in References 1 and 2; this article does not assert that none exist |
| RoboRXN's status in 2026, a Geodite paper, the next model release date, and any plan for a paid service | Not yet confirmed | Could not be confirmed; this article does not guess |
| Reading: materials AI is mainly open publication of research, not a product | Inference | Our own interpretation from the release form and the Form 10-K |
| Reading: models are put to work through joint research with partner companies | Inference | Our own interpretation from the JSR agreement and WG4M |
| Reading: fine-tuning on small data is the practical focus | Inference | Our own interpretation from the blog and the electrolyte example |
| Reading: IBM aims to supply a base on which materials companies grow models from their own data | Inference | Our own interpretation from the electrolyte example |
| Risk: no visible path to monetization | Inference | Our own interpretation from the release form |
Last updated 22 October 2026 / Troy Technical
The figures in this article rest on IBM’s SEC filings, IBM Research’s official material, model cards, papers, and JSR’s official release. Market-research estimates and press-only figures are not used. Figures described as claims by the company or study team were not reproduced or verified here. Items marked “Inference” are our own reading of the primary sources and are not statements by the company.