Overview
| Name | MoLFormer-XL · aliases: MoLFormer-XL, MoLFormer-XL-both-10pct |
|---|---|
| Full name | The paper gives no expanded name |
| Ambiguous names (not auto-linked) | MoLFormer |
| Family | Not a family entry. The later generative variant from the same organisation has no site entity record, so no family entry |
| Organization | IBM Research |
| Organization type | Big-tech lab |
| Country | United States |
| Model class | Property prediction (molecule-string language model) |
| Domain | Small molecules |
| First release | · paper revision Basis: arXiv 2106.09553 v1 submitted Note: The site used the 2022 journal year as the release; first release follows the preprint v1 source |
| Other milestones |
|
| Latest update | · journal Basis: Journal publication Journal publication; the code repository's last commit is 2023-10-16 (checkpoint usage documentation). This is the latest sourced substantive event source The most recent sourced event, not a claim that nothing changed since. |
| Status | Maintained only · No substantive event after 2023-10, though it is still widely used as a baseline |
| Openness | Code: Open · Apache-2.0 Weights: partial · Apache-2.0 (model card) Code is open; the public weights are a subset model trained on 10% of each corpus, and the full large-model weights are not released |
| Benchmarks reported | 1 counted (self 1) |
| Wet-lab validation | None recorded |
| Unverified fields | † release_milestones.code_release, release_milestones.weights_release, background.org_profile |
| As of | 2026-11-16 |
What it does
Treats molecule strings as text, using a twelve-layer Transformer with linear attention and rotary position embeddings so that billion-scale pretraining is feasible, then fine-tunes on downstream property tasks; the paper argues the representation captures molecular geometry and structure. Unlike graph models such as MolE it needs only strings rather than graphs or conformations, and unlike Uni-Mol2 it does not depend on 3D conformations.
Benchmarks reported
Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 1 of 1 pairings are counted.
| Benchmark | Relation | Eval set / protocol | Source type | Note and evidence |
|---|---|---|---|---|
| 03 MoleculeNet (+OGB) | Developer report, official protocol | — | Self-reported | MoleculeNet scaffold/random splits · evidence |
Other evaluations (outside the 44)
- Geometry probe of the representation (Own task set) Self-reported: The authors use self-built analyses to argue the string representation carries geometric and structural information
- Corpus-scale controls (Own task set) Self-reported: Controls across pretraining corpus fractions; the public release corresponds to the 10% subset setting
Developers report: linear attention makes billion-scale string pretraining feasible, and after fine-tuning it beats the supervised graph-neural-network baselines of the time on public property tasks.
Independent evidence: No independent evidence found. No independent evaluation of the model was found, and task-group numbers in third-party papers are not board results
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Up to 1.1 billion molecule strings from two public compound libraries; the public version uses 10% of each source |
| b. Private or licensed data | No | source |
| c. Own wet-lab data used for training | No | |
| d. Wet lab used only for validation | No | The paper reports no wet-lab validation source |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: The developer is the research division of a large US technology company, centred in New York State; its careers page claims over three thousand researchers across twelve labs on six continents (the page is undated). A change of research director appears only in second-hand sources with inconsistent months and is not concluded here.
- Funding: Not disclosed
- Products and partnerships: The company's chemistry foundation-model line includes this model, a later decoder-only generative model and a target-specific generative model, plus a public molecular foundation-model site; on-site comparisons are MolE (graph) and Uni-Mol2 (3D).
- Related models: MolE, Uni-Mol2, MiniMol
Reading notes
- These are 2022 results that later foundation models use as a baseline, yet the full large-model weights are not public, so independent reproduction is hard.
- MoleculeNet datasets are small and split-sensitive, and studies report that most pretrained embeddings do not beat classical fingerprints.
- None of the 22 dataset pages of the official TDC ADMET board carries an entry for this model (checked 2026-11-16), so task-group numbers in third-party papers are not board results.
- The comparison baselines are weak (supervised graph networks only), so cross-paper comparisons need fingerprint baselines added.
Caveats
- The full weights are not public; only the 10% subset model is available.
- First dates for code and weights could not be retrieved this round and are marked unverified.
- All benchmark results are author-reported with no board entry.
- No substantive update after 2023-10.
Sources
- Journal paper 2022 (journal)
- Preprint (figures follow v3) (preprint)
- Code repository (repository)
- Public subset weights and model card (package registry)
- Another task-group leaderboard overview (used to rule out a pairing) (leaderboard)
Corrections
2 corrections
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.