AI for Bio · Model

Uni-Mol2

3D-conformation molecular pretraining model scaled to the billion-parameter range to study scaling

Small moleculesPlatform companyChinaMaintained only中文

Overview

NameUni-Mol2 · aliases: Uni-Mol2
Full nameThe paper uses the model name only; no expanded name
Ambiguous names (not auto-linked)Uni-Mol
FamilyNot a family entry. Each generation has its own record; the bare name without a digit is ambiguous and is never auto-linked
OrganizationDP Technology, Peking University, AI for Science Institute, Beijing
Organization typePlatform company
CountryChina
Model classProperty prediction (3D-conformation molecular pretraining)
DomainSmall molecules
First release · paper revision
Basis: arXiv 2406.14969 v1 submitted (v2 on 2024-07-01)
Note: The site previously gave only 2024-06 source
Other milestones
  • First public evidence: · arXiv v1 source
  • Preprint v1: · arXiv v1 source
  • Code release: † · Pull request adding the model codebase merged into the main repository source
  • Weights release: † · Weights repository created with several model sizes uploaded (84M up to the billion range) source
Latest update† partly verified · code release
Basis: Pull request adding the model code to the main repository
Code and several pretrained weight sizes landed in the main repository and the weights host (weights repo created 2024-10-17); the paper was also accepted at a 2024 machine-learning conference; commits through 2025-05 are maintenance. This is the latest sourced substantive event
Note: The earlier 2024-10-10 release-tag framing is corrected source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · Latest substantive event in 2024-10, with maintenance commits only afterwards
OpennessCode: Open · MIT
Weights: Open
The repository is MIT overall and the weight sizes are public on the host; no separate weights licence is listed
Benchmarks reported1 counted (modified protocol 1)
Wet-lab validationNone recorded
Unverified fields† latest_update, release_milestones.code_release, release_milestones.weights_release, open_status.weights_license
As of2026-11-16

What it does

Pretrains on molecules represented as 3D conformations and pushes both parameter and conformation scale, reporting how size and data relate to downstream performance; the paper positions itself as a scaling study of molecular pretraining. Compared with MiniMol and MolE it needs pregenerated conformations, is two orders of magnitude larger, and reports quantum-chemistry property tasks rather than drug ADMET; the sibling successor Uni-Mol3 targets reactions.

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.

BenchmarkRelationEval set / protocolSource typeNote and evidence
03 MoleculeNet (+OGB)Developer report, modified protocol—Self-reportedQM9 only, scaffold split †partly verified · evidence

Other evaluations (outside the 44)

  • Off-site quantum-chemistry property dataset (Public benchmark not among the 44) Self-reported: A property dataset of several thousand molecules compared with the predecessor, in settings with and without atom and bond features; developer-reported
  • Own scaling study (Own task set) Self-reported: A self-built comparison across model sizes and data volumes supporting the scaling claim, with tasks chosen by the authors

Developers report: scaling parameters and conformations together improves quantum-chemistry property tasks over the predecessor and yields a scaling relationship for molecular pretraining.

Independent evidence: No independent evidence found. No third-party evaluation of this model on a site benchmark was found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesAbout 19 million molecules from the predecessor plus a subset of a public compound library (about 884 million compounds), about 800 million conformations in total source
b. Private or licensed dataNot disclosed
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNoThe paper reports no wet-lab validation source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The developer is an AI-for-Science tooling and platform company founded in Beijing in 2018, with a research cloud platform and laboratory and drug-design platforms; collaborators are a Chinese university and a Beijing AI-for-science institute.
  • Funding: A Series C of over RMB 800 million (about US$114 million) was announced 2025-12-24, led mainly by industrial and municipal funds; earlier rounds appear in the same company's other records.
  • Products and partnerships: The model series is open-sourced under the same organisation's public repositories alongside platform and tooling products; this model's goal is to test the scaling relationship of molecular pretraining.
  • Related models: Uni-Mol3, MiniMol, MolE, KERMT

Reading notes

  1. The reported tasks are quantum-chemistry properties, a different track from the ADMET competitors on TDC, so the two should not be compared directly.
  2. The scaling conclusion rests on author-chosen tasks; the benefit of scale on small ADMET datasets has no independent verification.
  3. Conformations must be generated in advance, making inference costlier than graph or string models.
  4. The weights are public, making it one of the few large molecular pretraining models outsiders can evaluate themselves; only a single subset of MoleculeNet is reported.

Caveats

  1. Only one subset of the benchmark is reported; the other datasets are not, so coverage is limited.
  2. The latest substantive event is 2024-10 and the weights date relies on a host-side creation date.
  3. No separate weights licence is listed in the repository.
  4. No wet-lab validation and no third-party evaluation.

Sources

  1. Model paper, arXiv 2406.14969 (preprint)
  2. Paper v2 full text (preprint)
  3. Series code repository (repository)
  4. Weights hosting repository (package registry)

Corrections

3 corrections
FieldWasNowBasisFound
latest_update2024-10-10 (code and weights release tag)2024-10-18 (PR merging the model code into the main repository); weights repository created 2024-10-17; the releases list ends at v0.2.1 (2024-07-06)source2026-11-16
release_milestones.journal_onlineNo conference or journal version seenThe paper was accepted at a 2024 machine-learning conference (no journal version)source2026-11-16
first_release2024-062024-06-21(arXiv v1)source2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.