AI for Bio · Model

ProGen3

Sparse mixture-of-experts autoregressive protein language model focused on scaling laws and alignment with laboratory data

ProteinsAI-native biotechUnited StatesMaintained only中文

Overview

NameProGen3 · aliases: ProGen3, progen3-3b
Full nameThe name is the product name; the paper gives no expanded form
Ambiguous names (not auto-linked)ProGen
FamilyPart of the ProGen family
OrganizationProfluent
Organization typeAI-native biotech
CountryUnited States
Model classProtein language model (autoregressive)
DomainProteins (also: Molecule generation)
First release · preprint
Basis: Preprint v1
Note: Weights were uploaded to the model hub on 2025-03-20; per this round's rule a repository or hub creation date is not used as first release source
Other milestones
  • First public evidence: · Preprint v1 source
  • Preprint v1: · Preprint v1 source
  • Weights release: † · Weights (112M-3B) uploaded to the model hub source
Latest update · no later event
Basis: No new version found after the preprint
The last repository commit, on 2025-07-21, changed a build file and is not a new version; no journal publication or successor model was found (as of 2026-11-16) source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · Only build-file level changes in the last 12 months, no new version
OpennessCode: Partly open
Weights: Non-commercial
Small and mid sizes are downloadable from the model hub under non-commercial terms while the largest size is unreleased; the site's "weights not open" needs correcting, and licence text was not re-checked item by item
Benchmarks reported1 counted (board 1)
Wet-lab validation1 recorded campaign
Unverified fields† open_status.weights_license, open_status.code_license, wetlab_validation, training_data.c_own_wetlab_for_training.detail
As of2026-11-16

What it does

Generates protein sequences autoregressively with a sparse mixture-of-experts architecture, studies parameter and data scaling systematically, and then aligns on limited laboratory data, aiming at expressible and functional proteins. It divides work with the same developer's Profluent-E1, which encodes and scores rather than generates, and sits on the same line as generative protein language models such as ESM3.

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 results—BoardEntry on the official summary table · evidence

Other evaluations (outside the 44)

  • Own scaling study (Own task set) Self-reported: Compares generation diversity and expressibility across sizes on in-house data; a developer-defined task set
  • Alignment against laboratory data (Wet-lab validation) Self-reported: Checks alignment against measured fitness; the data and its scale are not public, so it cannot be re-checked externally

Developers report: Developers report that larger models generate a wider range of expressible, functional proteins and can be aligned with a small amount of laboratory data.

Independent evidence: The official zero-shot board lists several of its sizes, an organiser-maintained third-party record.

Training data

Group in Figure 7: Own wet-lab training data.

ClassValueDetail
a. Public databasesYesAn in-house protein atlas v1: 3.4 billion full-length proteins and 1.1 trillion amino-acid tokens from genomes and metagenomes, cleaned in house source
b. Private or licensed dataNot disclosed source
c. Own wet-lab data used for trainingYesAlignment used "limited laboratory data" on activity, expression, stability and binding; its size and provenance are undisclosed source
d. Wet lab used only for validationYesSeparate synthesis, expression and antibody-candidate binding and developability tests were run; counts undisclosed 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.

Expression and antibody-candidate tests on generated sequences (showcase page)
Targets — · Designs tested — · Hits — · Success — · Best affinity —

Denominator basis: The showcase page gives no design count or denominator
The developer also reports a rise in correlation between predictions and measured fitness after alignment; it is self-reported with no per-item data and is not independent evidence
Executor: Executor undisclosed

Self-reported · † · †unverified · source

Background

  • Organization: The developer is an AI-native protein and gene-editing company founded in California in 2022, reported at about 20 staff in 2024 with no current public headcount.
  • Funding: A $35M round in 2024-03 led by one venture firm and a $106M round on 2025-11-19 co-led by two investors bring the total to about $150M; no valuation is officially disclosed.
  • Products and partnerships: The product line covers this model, Profluent-E1 and an open gene-editing protein; on 2026-04-28 it signed a multi-programme recombinase collaboration with a large pharmaceutical company, with a disclosed milestone ceiling but an undisclosed upfront.
  • Related models: Profluent-E1, ESM3

Reading notes

  1. Its focus is generation while the board that lists it measures zero-shot scoring; the two are not the same thing.
  2. Adjacent sizes sit close together on that board, so the paper's scaling conclusion cannot be checked there.
  3. The scale of the alignment data is undisclosed, so the reported gain cannot be reproduced independently.
  4. Training used a privately curated large corpus, so overlap with evaluation sets cannot be checked externally.

Caveats

  1. The site stated weights were not open; in fact small and mid sizes are downloadable and only the largest is unreleased, so the statement must be split by size.
  2. The wet-lab work is described only on the showcase page without counts; it is treated as unverified and no figures are given.
  3. The largest size is not on the official board, so any standing claimed for it has no source.

Sources

  1. Preprint (preprint)
  2. Developer showcase page (official page)
  3. Model hub weights page (model hub)
  4. Code repository (repository)
  5. Official zero-shot summary table (leaderboard)

Corrections

2 corrections
FieldWasNowBasisFound
open_status.weightsWeights not open (licence / early API)Small and medium weights publicly downloadable on the model hub (non-commercial-type terms); the largest size is not releasedsource2026-11-16
latest_updateWording that there was no later eventChanged to "no later sourced event found (as of 2026-11-16)", noting that the 2025-07-21 commit only changed build filessource2026-11-16

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