AI for Bio · Model

OpenFold3

Consortium-built Apache-2.0 open reproduction of AlphaFold 3 with a federated private-data fine-tuning track

Structure & co-foldingNonprofit instituteUnited StatesActive中文

Overview

NameOpenFold3 · aliases: OpenFold3, OpenFold3-preview, OpenFold3-preview2, OF3
Full nameNo expanded name; official materials use OpenFold3
Ambiguous names (not auto-linked)OpenFold
FamilyNot a family entry. Single model line; preview, preview2 and 0.5.x are versions of one model
OrganizationOpenFold Consortium, Open Molecular Software Foundation, Columbia University, Lawrence Livermore National Laboratory, Seoul National University
Organization typeNonprofit institute
CountryUnited States
Model classCo-folding (protein / nucleic acid / small molecule)
DomainStructure & co-folding (also: Small molecules)
First release · code release
Basis: GitHub preview release, consortium press release the same day source
Other milestones
  • First public evidence: · GitHub preview release source
  • Code release: · Preview code public source
  • Weights release: † · Preview weights public on Hugging Face with the release source
  • Official announcement: · Consortium press release source
Latest update · code release
Basis: GitHub release 0.5.1
v0.5.1 patch: lower inference/validation memory (MSA/template features as compact indices) and deterministic inference, no new weights; OpenBind-0 weights belong to v0.5.0 (2026-08-21)
Note: Commits continue to 2026-11-16; preview2 was 2026-03-13 source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Releases and ongoing commits within 12 months
OpennessCode: Open · Apache-2.0
Weights: Open · Apache-2.0
Code and weights Apache-2.0, commercial use allowed; the 0.5.x default weights OpenBind-0 are also Apache-2.0
Benchmarks reported3 counted (board 1, self 1, retrospective 1)
Wet-lab validationNone recorded
Unverified fields† release_milestones.weights_release, training_data.b_private_licensed
As of2026-11-16

What it does

Open PyTorch reproduction of AlphaFold 3 for protein, nucleic-acid and small-molecule complexes, with reproducible training and distillation pipelines. Unlike Protenix it adds federated fine-tuning on pharma private data; unlike AlphaFold 3 it differs in licence and reproducibility. AQAffinity is built on it.

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. 3 of 3 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
31 FoldBenchOfficial leaderboard or resultsmaintainer leaderboard (preview)BoardMaintainer leaderboard (preview) · evidence
30 PLINDER / PINDER / Runs N' PosesDeveloper report, official protocolRuns N' Poses (protein-ligand)Self-reportedRuns N' Poses appears in a README bar chart; no numbers in text †partly verified · evidence
01 CASPRetrospective dataset useCASP16 protein and RNA monomers (post hoc)Third-partyCASP16 protein/RNA monomers, README figure only, post hoc †partly verified · evidence

Other evaluations (outside the 44)

  • AF3 antibody-antigen set (Public benchmark not among the 44) Self-reported: preview2 protocol, figures only
  • Federated fine-tuning holdout (Own task set) partner: Evaluation on 1,056 held-out private structures, self-reported by Apheris, not independently verified

Developers report: the preview is the only model matching AlphaFold 3 on CASP16 RNA monomers, under a commercially usable Apache-2.0 licence with reproducible training.

Independent evidence: The FoldBench maintainer leaderboard lists the preview, run by the board; federated fine-tuning gains lack independent verification.

Training data

Group in Figure 7: Private or licensed data.

ClassValueDetail
a. Public databasesYes300,000+ experimental structures plus 40M+ synthetic/distilled entries (incl. 13M MGnify) source
b. Private or licensed dataYesFederated fine-tuning by five pharma firms on 20,167 private structures, 1,056 held out (Apheris self-reported) source
c. Own wet-lab data used for trainingNot disclosedWhether OpenBind-0 uses newly measured data is not stated source
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: OpenFold is a non-profit AI research consortium hosted by the Open Molecular Software Foundation, with academic and industry members (NVIDIA cites 40+ organizations). OpenFold3 was co-developed by a US university lab, a US national laboratory and a South Korean university group.
  • Funding: Members contribute funding, compute and expertise; public materials name no specific grantor.
  • Products and partnerships: The AISB network, run by Apheris, has AbbVie, Astex, BMS, J&J and Takeda continue training via federated learning (source lists differ slightly); Apheris also offers ApherisFold; NVIDIA ships a NIM package; default weights OpenBind-0 were trained by the UK OpenBind project with a 2025-06 cutoff.
  • Related models: AlphaFold 3, Protenix, RF3, Boltz-1 / Boltz-1x, AQAffinity

Reading notes

  1. Its most informative feature is the federated route: whether private data that never leaves the owner improves a general model; the only evidence so far is Apheris's own account.
  2. The developers' own evaluations are figures without numbers and cannot be compared directly with other models.
  3. The 2026-10-28 release 0.5.1 is only a patch; new weights belong to 0.5.0 (2026-08-21); do not conflate them.
  4. Apache-2.0 plus reproducible training gives it distinct value where self fine-tuning is needed, independent of leaderboard rank.

Caveats

  1. No PoseBusters result; it is not in the official evaluation set.
  2. Sources disagree slightly on the list of federated pharma partners.
  3. The CASP16 pairing is post-hoc dataset reuse, not participation.

Sources

  1. OpenFold3 GitHub repository (repository)
  2. OpenFold3 GitHub releases (repository)
  3. OpenFold Consortium press release (2025-10-28) (press)
  4. OpenFold3-preview2 technical report (preprint)
  5. OpenFold3 on Hugging Face (repository)
  6. NVIDIA OpenFold3 model card (official site)
  7. AISB federated fine-tuning report (press)
  8. OpenBind third-party platform page (second-hand) (second-hand) second-hand

Corrections

3 corrections
FieldWasNowBasisFound
latest_update2026-08-212026-10-28 (tag 0.5.1)source2026-11-16
latest_update.descriptionv0.5.0 OpenBind model release: OpenBind-0 weights and recipesv0.5.1 patch: memory reduction and deterministic inference, no new weights; OpenBind-0 weights belong to v0.5.0 (2026-08-21)source2026-11-16
latest_update.event_kindnew-weightscode-releasesource2026-11-16

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