AI for Bio · Model

ESM3

Multimodal generative protein language model; sequence, structure and function prompt each other

ProteinsAI-native biotechUnited StatesMaintained only中文

Overview

NameESM3 · aliases: ESM3, esm3-sm-open-v1, ESM-3
Full nameOfficial materials use ESM3 only, with no expanded name
Ambiguous names (not auto-linked)ESM, ESM2
FamilyNot a family entry. The open small model and the larger API version are size variants of one model, not a family
OrganizationEvolutionary Scale, Chan Zuckerberg Biohub
Organization typeAI-native biotech
CountryUnited States
Model classpLM / design (multimodal generation)
DomainProteins (also: Molecule generation, Structure & co-folding)
First release · new weights
Basis: Developer ESM3 release blog, with esm3-sm-open-v1 weights posted on a model host the same day source
Other milestones
  • First public evidence: · Release blog and open weights, the earliest provable public evidence source
  • Preprint v1: · bioRxiv preprint (2024.07.01.600583) source
  • Journal online: † · Science online (issue 2025-02-21) source
  • Weights release: · Open small-model weights posted with the release blog source
  • Official announcement: · Developer release blog source
Latest update† partly verified · journal
Basis: Science publication (online 2025-01, issue 2025-02-21) plus the API public beta
Note: Day-level date unverified; see the journal_online note source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · The SDK is updated through v3.4.x, with no new ESM3 weights
OpennessCode: Open · SDK on a public repository
Weights: Research use only · open small model (esm3-sm-open-v1) under non-commercial terms; larger models via API
Small-model weights are downloadable while larger models come only through the API and partners; the two tiers differ and citations must say which was used
Benchmarks reported4 counted (board 3, retrospective 1)
Wet-lab validation1 recorded campaign
Unverified fields† latest_update.date, release_milestones.journal_online.date, benchmarks_reported[2]
As of2026-11-16

What it does

Tokenizes sequence, three-dimensional structure and function annotations and runs masked generation in one transformer, so any modality can prompt any other; the pitch is programmable protein design. That is a different job from its representation-only sibling ESM C.

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. 4 of 4 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
02 ProteinGymOfficial leaderboard or resultsDMS substitutions zero-shot, maintainer-run summary table (updated 2025-06-13); open 1.4B entryBoardEntry on the official summary table · evidence
34 AbBiBenchOfficial leaderboard or resultsbenchmark-registered entryBoardBenchmark-registered score †unverified · no source link
40 MotifBenchOfficial leaderboard or resultsMotifBench submission leaderboard (2025-05 submission)BoardEntry on the MotifBench submission board (11.38, 2025-05). · evidence
01 CASPRetrospective dataset useCASP14 (71 targets) and CASP15 (70 targets) time-held-out structure and contact predictionThird-partyCASP14/15 time-split post hoc · evidence

Other evaluations (outside the 44)

  • esmGFP design validation (Wet-lab validation) Self-reported: Wet-lab testing of two rounds of 96 designs; the developers' flagship case, outside the site's 44 benchmarks
  • Viral-protein weakness noted by maintainers (third party commentary) Third-party: Benchmark maintainers note that the ESM line degrades markedly on viral proteins and that zero-shot ability plateaus beyond roughly 1B-4B parameters

Developers report: prompting across all three modalities yields genuinely new functional proteins, demonstrated by esmGFP at about 58% sequence identity to the nearest known fluorescent protein (self-reported).

Independent evidence: The ProteinGym and MotifBench entries were run by the organizers; no independent replication of the wet lab was found.

Training data

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

ClassValueDetail
a. Public databasesYesUniRef, MGnify, JGI, OAS, PDB, AFDB and others; 2.78B natural proteins for the largest model, or 3.15B sequences, 236M structures and 539M function-annotated proteins once predicted and synthetic data are included source
b. Private or licensed dataNot disclosedAll listed sources are public databases; no licensed private data disclosed source
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationYesThe esmGFP designs were wet-lab tested, with no claim of feeding results back into training 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.

esmGFP fluorescent-protein design (developer blog)
Targets — · Designs tested 96 · Hits — · Success — · Best affinity —

Denominator basis: Two rounds of 96 generated designs each were tested; the blog gives no hit count, so no hit rate is stated
The site said 'synthesized' where it should say 'tested'; round two produced designs as bright as the natural protein. Developer-run, not an independent replication.

Self-reported · · verified · source

Background

  • Organization: Evolutionary Scale was founded in 2023-07 by the former Meta FAIR ESM team (which left in 2023-04) as a public benefit corporation, about 20 people at funding; in 2025-11 the team joined Chan Zuckerberg Biohub on undisclosed terms.
  • Funding: Seed round of $142M on 2024-06-25 led by Lux Capital with two angel investors, joined by Amazon and NVentures (NVIDIA's venture arm); valuation undisclosed.
  • Products and partnerships: Siblings are ESM C and ESMFold2, with access also via API; the Biohub that absorbed the team says it will scale to 10,000 GPUs by 2028. Competing lines include ProGen3 and Profluent-E1.
  • Related models: ESM C, ESMFold2, ProGen3, Profluent-E1

Reading notes

  1. The board entry is the maintainer-run open 1.4B model, not the larger API model; cite the summary table's date.
  2. esmGFP shows a functional protein can be generated, but one case is not a hit rate, and the blog gives no hit count.
  3. The paper reports only post-hoc CASP14/15 time splits; the claim of better CASP16 results has no primary source and should not be cited.
  4. The AbBiBench row rests on the site's own page and should be read as unverified.

Caveats

  1. The Science online date is day-level but unverified; only a created date and the print issue are checkable.
  2. The two rounds of 96 esmGFP designs give a denominator but no hit count, so any hit rate is extrapolation.
  3. The open small model and the API-only large model differ in licence and do not stand in for each other.
  4. Per this round's version guard, 'ESM2' must not link to this record.

Sources

  1. ESM3 release blog (company blog)
  2. ESM3 release note (weights posted) (company blog) ·
  3. ESM3 journal paper (journal) · †
  4. ESM3 preprint full text (preprint)
  5. ProteinGym official summary table (leaderboard)
  6. MotifBench repository and leaderboard (leaderboard)
  7. esm SDK releases (repository)
  8. Seed-round report (press)
  9. Announcement of the team joining Biohub (press)

Corrections

2 corrections
FieldWasNowBasisFound
wetlab_validation[0]Two rounds of 96 designs each 'synthesized'Two rounds of 96 designs each 'tested'; the blog gives no hit count, so no hit ratesource2026-11-16
latest_updateThe Science online date was treated as confirmedThe day-level online date is marked unverified; only a created date and the print issue are checkablesource2026-11-16

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