AI for Bio · Model

Uni-Mol3

Multi-molecular reaction foundation model covering products, retrosynthesis, conditions and yield

Small moleculesPlatform companyChinaActive中文

Overview

NameUni-Mol3 · aliases: Uni-Mol3
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
Organization typePlatform company
CountryChina
Model classReaction model / retrosynthesis (multi-molecular foundation model)
DomainSmall molecules (also: Molecule generation)
First release · paper revision
Basis: arXiv 2508.00920 v1 submitted (v2 on 2025-08-12) source
Other milestones
  • First public evidence: † · Initial code commit (v1.0), earlier than the paper's v1 source
  • Preprint v1: · arXiv v1 source
  • Code release: † · Initial repository code v1.0 source
  • Weights release: † · Repository commit adding weight download links source
Latest update† partly verified · code release
Basis: Repository commit adding weight download links
A repository commit added download links for pretrained and fine-tuned weights; initial code dates to 2025-07-18 (v1.0). No journal version was found. This is the latest sourced event
Note: The nature of the event is inferred from the commit message source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Substantive repository commit within 12 months
OpennessCode: Open · MIT
Weights: Open · MIT (per repository)
Code and weights are MIT per the repository; the licensed patent-reaction datasets cannot be redistributed and only preprocessing scripts are provided
Benchmarks reported1 counted (self 1)
Wet-lab validationNone recorded
Unverified fields† latest_update, release_milestones.code_release, release_milestones.weights_release, release_milestones.first_public_evidence, training_data.a_public
As of2026-11-16

What it does

Treats molecules and reactions as multi-molecular systems, pretraining on molecules and then on reactions at scale, with one model covering forward product prediction, retrosynthesis, condition recommendation and yield prediction. Unlike the reasoning-style retrosynthesis of RetroDFM-R it uses representation pretraining plus reaction fine-tuning; sequence models such as Chemformer are its main retrosynthesis comparison, and Uni-Mol2 is the sibling predecessor.

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
07 USPTO family (+PaRoutes)Developer report, official protocol—Self-reportedUSPTO-MIT/50K/Condition · evidence

Other evaluations (outside the 44)

  • Retrosynthesis test set inside the licensed database (Own task set) Self-reported: Developer-reported retrosynthesis on a test split of the licensed patent-reaction database; outsiders cannot obtain the data, so it is not reproducible
  • Reaction condition and yield tasks (Own task set) Self-reported: Self-built multi-task evaluation covering condition recommendation and yield prediction

Developers report: one model unifies product prediction, retrosynthesis, conditions and yield, with molecule pretraining plus large-scale reaction pretraining giving broadly usable performance.

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: Private or licensed data.

ClassValueDetail
a. Public databasesYesAbout 19 million molecules for pretraining from public compound and bioactivity libraries; the paper's description of the sources is internally inconsistent source
b. Private or licensed dataYesA commercially licensed patent-reaction database, about 16.67 million records reduced to 11,973,789 after cleaning, not redistributable under the licence source
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 platform company founded in Beijing in 2018, working across drug design, materials and battery research; the collaborator is a mathematics and machine-learning unit at a Chinese university. No citable headcount exists in public material, so none is given.
  • Funding: Series A of tens of millions of dollars in 2021-08; Series B of tens of millions in 2022-01; over RMB 700 million in 2023-08; a Series C of over RMB 800 million (about US$110 million) on 2025-12-24 with industrial and municipal funds. Outlets name the middle rounds inconsistently, so only years and amounts are recorded.
  • Products and partnerships: Products include a research cloud platform, a drug-design platform and docking and free-energy tooling; the model series is open-sourced in public repositories. This model's reaction pretraining depends on commercially licensed data that outsiders cannot reproduce.
  • Related models: Uni-Mol2, Chemformer, RetroDFM-R

Reading notes

  1. Its self-reported retrosynthesis result sits below Chemformer in the same table, so the selling point is multi-task unification rather than a single best score (no figures are transcribed here).
  2. Reaction pretraining uses a commercially licensed database that outsiders cannot reproduce; only code and weights are open.
  3. The USPTO test sets suffer template leakage, which limits what high scores mean.
  4. The nature of the latest update is inferred from a commit message; the weights and code axes should be read separately.

Caveats

  1. All benchmark results are developer-reported; there is no board entry and no third-party evaluation.
  2. The paper's description of the public portion of the training data is internally inconsistent.
  3. The licensed reaction database cannot be redistributed, so pretraining is not reproducible.
  4. Both the weights-link event and the journal status are unverified.

Sources

  1. Model paper, arXiv 2508.00920 (preprint)
  2. Paper v2 full text (preprint)
  3. Code and weights repository (repository)

Corrections

None recorded.

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