AI for Bio · Model

ESMFold2

Protein-language-model co-folding that needs no MSA retrieval, released alongside a sequence atlas

Structure & co-foldingNonprofit instituteUnited StatesActive中文

Overview

NameESMFold2 · aliases: ESMFold2, ESMFold2-Fast, ESMFold2 Pro
Full nameOfficial materials use ESMFold2 only; no expanded name
Ambiguous names (not auto-linked)ESMFold, ESM
FamilyNot a family entry. Covers ESMFold2 with its Fast and Pro variants; ESM C and the ESM Atlas are separate entries
OrganizationBiohub, EvolutionaryScale team
Organization typeNonprofit institute
CountryUnited States
Model classCo-folding (single-sequence PLM)
DomainStructure & co-folding (also: Proteins)
First release · product update
Basis: Research-institute release page (HF repos created 2026-05-13, not used as first release) source
Other milestones
  • First public evidence: · Institute release page source
  • Weights release: · Weights public on Hugging Face at release source
  • Official announcement: · Institute release page, published with ESM C and the ESM Atlas source
Latest update · patch
Basis: esm SDK v3.4.3
v3.4.3 is bug fixes plus a transformers dependency bump; the new weights belong to the previous release v3.4.2 (2026-10-30, with the 2025-cutoff weights and the Pro folding API), and HF weights were updated 2026-09-14
Note: Read the SDK and weights axes separately source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · The SDK still ships patch releases
OpennessCode: Unclear
Weights: Open
Weights are public on Hugging Face, but licence metadata is inconsistent across the organisation's model cards, so neither term is settled
Benchmarks reported2 counted (board 1, self 1)
Wet-lab validation1 recorded campaign
Unverified fields† wetlab_validation[0].success_rate, benchmarks_reported[1].source_url, open_status.code
As of2026-11-16

What it does

Puts evolutionary information inside a pretrained protein language model, predicts multi-chain complexes without MSA retrieval at inference, and ships a binder design pipeline. Unlike AlphaFold3 and Boltz-2 it needs no MSA and runs faster; ESM3 and ESM C are its siblings.

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. 2 of 2 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
34 AbBiBenchOfficial leaderboard or resultsAbBiBench structure-prediction entryBoardBenchmark-registered ipSAE entry †unverified · no source link
31 FoldBenchDeveloper report, official protocolFoldBench no-MSA complexes / antibody-antigenSelf-reportedModel-card figures are second-hand; another paper reports figures on its own antibody subset †partly verified · evidence

Other evaluations (outside the 44)

  • Sequence and structure atlas released alongside (resource release) Self-reported: The release includes an atlas of 6.8B sequences and 1.1B predicted structures; these are official scale figures, not benchmark results
  • Competitor-run evaluation (competitor reported) Third-party: The OpenDDE report evaluates this model on a self-defined antibody subset under a different protocol, with no independent replication
  • Hosting and second-hand accounts (second hand) Third-party: A third-party platform hosts the model, and its card figures mostly reach readers through platform and community write-ups (second-hand)

Developers report: Developers report matching or beating AlphaFold3 and Protenix-v1 on protein-protein and antibody-antigen targets without MSAs (self-reported, no independent replication).

Independent evidence: No independent evidence found. Both figure sets are self-reported, by developer and competitor; neither was replicated

Training data

Group in Figure 7: Public data + wet lab for validation only.

ClassValueDetail
a. Public databasesYesPDB plus AFDB (cut-off 2021-09), with the ESM C sequence corpus for the language model source
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNoNo claim that wet-lab results were fed back into training
d. Wet lab used only for validationYesBinder designs for five targets were assayed in the wet lab (unverified) source

Wet-lab validation

Every rate carries its numerator and denominator. Claims that could not be verified are shown as unverified, without numbers; different paper versions are separate entries.

Binder campaign shared across the release (institutional news plus shared preprint)
Targets — · Designs tested — · Hits — · Success — · Best affinity —

Denominator basis: Denominators undisclosed; the preprint full text is rate-limited and the campaign is shared across several models in one release, so it cannot be attributed to a single model (unverified)
Not an ESMFold2-specific wet-lab paper and not an independent replication; no hit-rate figures are given (unverified)

Self-reported · · †unverified · source

Background

  • Organization: Biohub is a philanthropically funded nonprofit medical research organization that consolidated related biomedical research in 2025-11; the EvolutionaryScale team joined on 2025-11-06 in a strategic transaction with undisclosed terms.
  • Funding: Nonprofit funded by its philanthropic backer; amounts undisclosed
  • Products and partnerships: Released 2026-05-27 together with ESM C and the atlas and hosted free on the organisation's own platform; the developers also ran the pipeline to design binders for five cancer and immunology targets.
  • Related models: ESM3, ESM C, AlphaFold 3, Boltz-2

Reading notes

  1. Self-reported numbers sit on FoldBench and are not comparable with PoseBusters-style protocols.
  2. The antibody-antigen claim exists in two different protocols, one from the developer and one from a competitor.
  3. The real value of dropping MSAs is speed and ease of use; the accuracy claim still awaits independent evaluation.
  4. The design-validation denominators are undisclosed, so hit rates cannot be cited.

Caveats

  1. The wet-lab campaign is shared across the release's models and the preprint is rate-limited, so no figures are given (unverified).
  2. The Runs N' Poses claim has no matching figure in official material and is treated as unconfirmed.
  3. Licence metadata is inconsistent across the organisation's models, so openness cannot be settled.
  4. The AbBiBench line rests on the site's own page and is shown as unverified.

Sources

  1. Institute release page: a world model of protein biology (official site) ·
  2. Hugging Face model card (repository)
  3. esm SDK releases (repository)
  4. Community write-up (second-hand) (second-hand) second-hand
  5. Announcement of the team joining (press)

Corrections

3 corrections
FieldWasNowBasisFound
latest_update2026-10-302026-11-09 (v3.4.3 is the later release)source2026-11-16
latest_update.event_kindnew weightsPatch: v3.4.3 is only bug and dependency fixes; the new weights belong to the 2026-10-30 releasesource2026-11-16
wetlab_validationhit-rate ranges for minibinders and antibody-derived formatsNo figures: the preprint is rate-limited and the campaign is shared across the release's models (unverified)source2026-11-16

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