Overview
| Name | MiniMol · aliases: MiniMol, minimol |
|---|---|
| Full name | The paper uses the model name only; no expanded name |
| Ambiguous names (not auto-linked) | None |
| Family | Not a family entry. A single model; the group's earlier MolE has its own record |
| Organization | Valence Labs (Recursion Pharmaceuticals) |
| Organization type | AI-native biotech |
| Country | United States |
| Model class | Property prediction (molecular foundation model, graph model) |
| Domain | Small molecules |
| First release | · paper revision Basis: arXiv 2404.14986 v1 submission date (also accepted at a machine-learning conference workshop) source |
| Other milestones | |
| Latest update | · code release Basis: Public package release 1.3.5 on the registry Public package releases 1.3.4 (2024-08-05) and 1.3.5 (2024-08-22); the matching repository exists with no releases or tags and a last commit of 2025-05-29 for storage-path maintenance. This is the latest sourced event Note: Corrects the site's "no update since release" wording: these package versions post-date the first release source The most recent sourced event, not a claim that nothing changed since. |
| Status | Maintained only · Maintenance commits only, the last on 2025-05-29 |
| Openness | Code: Open Weights: Unclear The authors call it open source and a public package is installable; this round the official repository and weights were not found under the expected organisation path (the old path is dead), so licence and weight status are unverified |
| Benchmarks reported | 1 counted (board 1) |
| Wet-lab validation | None recorded |
| Unverified fields | † open_status.code, open_status.weights, open_status.code_license, open_status.weights_license, training_data.b_private_licensed |
| As of | 2026-11-16 |
What it does
Takes the small-model, many-supervised-tasks route: about ten million parameters pretrained on millions of molecules, hundreds of millions of labels and roughly three thousand sparsely defined graph- and node-level tasks spanning quantum chemistry and biology, then applied to downstream property prediction. Against the group's earlier MolE it uses a smaller molecule set with denser labels, and against Uni-Mol2 it is more than two orders of magnitude smaller, emphasising parameter efficiency.
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 |
|---|---|---|---|---|
| 04 TDC (ADMET + DTI) | Official leaderboard or results | — | Board | Self paper plus official board entries · evidence |
Other evaluations (outside the 44)
- Parameter-efficiency comparison (Comparison against baselines only) Self-reported: Uses a far smaller parameter count than peer foundation models as its selling point and comparison; a self-reported framing, not an independent evaluation
Developers report: dense multitask supervised pretraining lets a ten-million-parameter model significantly beat the previous best molecular foundation model on most ADMET tasks.
Independent evidence: The public per-task ADMET board carries its entries so they can be read alongside other submissions, though the organisers did not re-run the model independently.
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | About 6 million molecules, 500 million labels and 3,300 graph- and node-level tasks; the abstract lists no database names source |
| b. Private or licensed data | Not disclosed | |
| c. Own wet-lab data used for training | No | |
| d. Wet lab used only for validation | No |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: The developer is the AI research arm of a listed AI-pharma company, announced in 2023 at a machine-learning conference and formed by merging the parent's machine-learning team with an acquired molecular-design company; the parent was founded 2013-11-05 and is based in Salt Lake City, Utah.
- Funding: Not disclosed
- Products and partnerships: The research arm co-developed the co-folding and affinity model Boltz-2 with a US university team; its earlier in-house model is MolE.
- Related models: MolE, Boltz-2, Uni-Mol2, KERMT
Reading notes
- "Better on 17 tasks" is a comparison with the previous best foundation model, not with all methods: on the official per-task board several fingerprint and gradient-boosting models rank higher.
- It is a workshop paper with no journal version and only a v1 preprint.
- The benchmark's datasets are small with high variance, so rank gaps are close to noise.
- The official repository and weights were not located this round, so openness is treated as unverified.
Caveats
- Open-source and weight status unverified (the expected repository path is dead).
- The abstract does not name the foundation model it is compared against.
- Workshop paper only, no journal version.
- The site's "no update since release" wording is corrected to the public package releases.
Sources
- Preprint arXiv 2404.14986 (preprint)
- Conference workshop entry (official site)
- Package registry page (package registry)
- Code repository (no releases or tags) (repository)
- Public ADMET per-task board (sample task) (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.