AI for Bio · Model

ESM C

Sequence-only protein language model positioned as a replacement for ESM-2

ProteinsAI-native biotechUnited StatesActive中文

Overview

NameESM C · aliases: ESM C, ESMC, ESM Cambrian, esmc-300m, esmc-600m, ESMC 6B
Full nameESM Cambrian
Ambiguous names (not auto-linked)ESM, ESM2
FamilyNot a family entry. 300M/600M/6B are size variants of one model and the 2026-05 refresh is recorded as a milestone and note, not a family
OrganizationEvolutionary Scale, Chan Zuckerberg Biohub
Organization typeAI-native biotech
CountryUnited States
Model classpLM (representation)
DomainProteins
First release · new weights
Basis: ESM Cambrian blog, with the 300M/600M weights first committed to a model host the same day source
Other milestones
  • First public evidence: · Blog plus the 300M/600M weights, the earliest provable public evidence source
  • Preprint v1: † · Protein world-model preprint released with the Biohub-era generation source
  • Weights release: · Weights axis: updated ESMC weights and SAE models; the 6B model opened separately on 2025-12-01 and does not belong to this event source
  • Official announcement: · The institute announced ESMC, ESMFold2 and a sequence atlas as a world model of protein biology source
Latest update · code release
Basis: esm SDK v3.4.0 added full ESMC support (SDK and weights are two separate axes)
Note: Later SDK releases v3.4.1-v3.4.3 run to 2026-11-09; the latest weights-axis event is 2026-05-27 source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Both the SDK and weights axes had events within 12 months
OpennessCode: Open · SDK on a public repository
Weights: Open · MIT plus third-party notice (per current blog and model card)
The current blog and model card state that 300M/600M/6B are all open weights under MIT for academic and commercial use; check the third-party notice first
Benchmarks reported1 counted (board 1)
Wet-lab validation1 recorded campaign
Unverified fields† wetlab_validation[0], release_milestones.paper_v1, training_data.a_public
As of2026-11-16

What it does

A sequence-only representation model for embeddings, mutation-effect scoring and downstream supervised tasks, a different job from the multimodal generation line of ESM3. The updated 6B version has 80 layers and is frozen to serve as the backbone of ESMFold2.

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
02 ProteinGymOfficial leaderboard or resultsDMS substitutions zero-shot, maintainer-run summary table (updated 2025-06-13); 300M and 600M entriesBoardEntry on the official summary table · evidence

Other evaluations (outside the 44)

  • Size-matched comparison against ESM-2 (Comparison against baselines only) Self-reported: The official blog says the 300M model matches ESM-2 650M and the 600M model compares with ESM-2 3B (developer self-report)
  • Binder design wet lab (unverified) (Wet-lab validation) Self-reported: Five-target binder experiments in the updated release; total design counts undisclosed, so no figures
  • Structure boards entered via a downstream model (related model evaluation) Self-reported: On structure boards the entrant is ESMFold2, built on the frozen 6B; ESM C itself does not report there

Developers report: better representation quality per unit cost than ESM-2, with 300M matching ESM-2 650M and 600M matching ESM-2 3B (self-reported, no third-party check found).

Independent evidence: Both ProteinGym entries were run by the benchmark maintainers; no independent evidence exists for the wet-lab side.

Training data

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

ClassValueDetail
a. Public databasesYesUniRef, MGnify and JGI clustered at 70% identity into about 83M, 372M and 2B clusters; the institutional page separately states about 2.8B training sequences (developer framing) source
b. Private or licensed dataNot disclosedThe official pages list only those public databases and disclose no private data source
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationYesFive-target binder experiments in the updated release, with no claim of feeding results back into training; denominators and attribution 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.

Five-target binder validation (unverified) (institutional news plus shared preprint)
Targets — · Designs tested — · Hits — · Success — · Best affinity —

Denominator basis: Unverified: denominators, the hit definition and whether rates are per design or per target are all undisclosed and the preprint full text was rate-limited, so no figures are given
Unverified: the campaign is shared across several models in the same release and is not an ESM C-specific result; the two shared rate ranges also attributed to ESMFold2 are not adopted this round.

Self-reported · † · †unverified · source

Background

  • Organization: Evolutionary Scale was founded in 2023-07 by the former Meta FAIR ESM team as a public benefit corporation; in 2025-11 the team joined Chan Zuckerberg Biohub on undisclosed terms and the model repositories moved to that organization.
  • Funding: Seed round of $142M on 2024-06-25 led by Lux Capital with two angel investors, joined by Amazon and NVentures; valuation undisclosed, as are the Biohub transaction terms.
  • Products and partnerships: Siblings are ESM3 and ESMFold2; on 2026-05-27 ESMC, ESMFold2 and a sequence atlas were announced together as a world model. Competing lines include ProGen3 and Profluent-E1.
  • Related models: ESM3, ESMFold2, ProGen3, Profluent-E1

Reading notes

  1. The 300M and 600M entries sit almost level on the board: doubling size buys nothing on zero-shot mutation effects.
  2. The SDK and the weights are two version axes and should be read separately; the 6B model opened on 2025-12-01, not in the 2026-05-27 event.
  3. Licensing follows the current blog and model card (MIT plus a third-party notice); check that notice before use.
  4. The binder rates are developer framing from a campaign shared across models; no figures are given this round and they cannot be compared with design models.

Caveats

  1. The two binder rate ranges shared with ESMFold2 are unverifiable; the preprint full text was rate-limited.
  2. The latest update is an SDK event, not new weights; the older record on the site conflated the two.
  3. The figure of about 2.8B sequences appears only on the institutional page, uses a different basis from the cluster counts and must not be added to them.
  4. Per this round's version guard, 'ESM2' must not link to this record.

Sources

  1. ESM Cambrian official blog (company blog) ·
  2. ESM C 600M model card (official site)
  3. Updated ESMC 6B model card (official site)
  4. Institutional world-model release page (official site) ·
  5. Protein world-model preprint (full text not retrieved) (preprint)
  6. esm SDK releases (repository) ·
  7. ProteinGym official summary table (leaderboard)

Corrections

5 corrections
FieldWasNowBasisFound
latest_update2026-05-27 (updated weights)2026-08-27 (SDK v3.4.0 added full ESMC support)source2026-11-16
latest_update.event_kindnew weightscode release; the SDK and weights are recorded as two axessource2026-11-16
release_milestones.weights_releaseThe 6B model opened with the 2026-05-27 updateThe 6B model opened on 2025-12-01 and is not part of the 2026-05-27 eventsource2026-11-16
wetlab_validation[0]The two binder rate ranges on five targets were published as confirmed figuresNo figures: the full text was rate-limited and the campaign is shared across models, so it is marked unverifiedsource2026-11-16
open_statusEarly framing: 6B via API only, with non-commercial terms for 300M/600MThe current blog and model card state all three sizes are open weights under MIT plus a third-party noticesource2026-11-16

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