AI for Bio · Model

BoltzGen

Open general-purpose binder design model built on a co-folding architecture

ProteinsOpen-source communityUnited StatesActive中文

Overview

NameBoltzGen · aliases: BoltzGen
Full nameNo expanded name; the release page uses BoltzGen only
Ambiguous names (not auto-linked)Boltz
FamilyNot a family entry. Record covers BoltzGen only; the closed successor BoltzProt-1 is unverified and not merged in
OrganizationMIT Jameel Clinic, Boltz PBC
Organization typeOpen-source community
CountryUnited States
Model classDesign (co-folding generation: protein / peptide / nanobody)
DomainProteins (also: Molecule generation, Structure & co-folding)
First release† partly verified · code release
Basis: First public code and weights (first repository commit, corroborated by two technology outlets)
Note: The site said 2025-11; the manuscript sits on the release page and repository, the bioRxiv entry (2025-11-24) is later source
Other milestones
  • First public evidence: † · First repository commit, open-source release source
  • Preprint v1: † · bioRxiv v1 (10.1101/2025.11.20.689494), not journal-published source
  • Code release: † · Code public on GitHub source
  • Weights release: † · Weights released with the code source
  • Official announcement: † · Research-institute release page source
Latest update · new version
Basis: GitHub release v0.3.2
v0.3.2 fixes binding_types; v0.3.1 (2026-03-06) added per-residue constraints and ipSAE, v0.3.0 (2026-02-10) symmetric inverse folding and covalent bonds; commits continue to 2026-11-01 source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Version releases and ongoing commits within 12 months
OpennessCode: Open
Weights: Open
Code and weights both public on GitHub; licence text not retrieved this round, so both terms are left empty
Benchmarks reported0 counted · none of the 44
Wet-lab validation5 recorded campaigns
Unverified fields† wetlab_validation[4], background.funding, background.products_partnerships, first_release, open_status.code_license, open_status.weights_license
As of2026-11-16

What it does

Puts structure prediction and design in one generative framework, generating binders conditioned on protein, peptide or small-molecule targets under design constraints. Shares lineage with Boltz-2 but is not a structure predictor; unlike the backpropagation route of BindCraft it is a trained generative model, and unlike the closed Latent-X/X2 and Chai-2 its weights are open and reproducible.

Benchmarks reported

None of the 44 benchmarks. This model is on the zero-pairing list.

How it evaluates itself: Novel-target wet-lab benchmark; Known-target check; Partner laboratory campaigns; Folding-model parity

Other evaluations (outside the 44)

  • Own novel-target wet-lab evaluation (Wet-lab validation) Self-reported: Ten novel targets, two modalities, top 15 designs per modality assayed as the developers' own evaluation set
  • Known-target check (Wet-lab validation) Self-reported: Reuses another model's five published easy targets as a control
  • Partner laboratory campaigns (Case study) Self-reported: Functional and affinity readouts across campaigns; scale and executor attribution unverified
  • Parity with folding models (Comparison against baselines only) Self-reported: The release page claims parity with contemporary folding models and gives no metrics

Developers report: one open model designs protein, nanobody and peptide binders across target types and reaches nanomolar binders on novel targets without known binders.

Independent evidence: No independent evidence found. Wet-lab work was CRO- and partner-executed but compiled by the developers; no independent replication or third-party evaluation found

Training data

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

ClassValueDetail
a. Public databasesYesPDB, AFDB and similar public structure data (Boltz family) source
b. Private or licensed dataNoNone for this model; a later pharma collaboration uses that company's historical data, outside this model source
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationYesDesigns assayed in the wet lab, run by a CRO with orthogonal confirmation by a second vendor author-hosted manuscript (link not published)

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.

Novel hard targets, nanobody arm (manuscript (paper v2 lineage))
Targets 10 · Designs tested 15 · Hits 6 · Success 6/10 · Best affinity —

Denominator basis: Ten novel targets with no bound structure in the PDB (1%-27.9% identity to any protein seen bound); 60,000 nanobodies designed per target, top 15 per modality assayed
Denominator follows the manuscript (10), not the release page (9)
Executor: CRO-executed with orthogonal confirmation by a second vendor; not an independent evaluation

Self-reported · † · verified · author-hosted manuscript (link not published)
Novel hard targets, miniprotein arm (manuscript (paper v2 lineage))
Targets 10 · Designs tested 15 · Hits 5 · Success 5/10 · Best affinity —

Denominator basis: Same ten targets; 60,000 miniproteins designed per target, top 15 assayed
A second 6/9 figure once published had no source and is removed
Executor: CRO-executed with orthogonal confirmation by a second vendor

Self-reported · † · verified · author-hosted manuscript (link not published)
Developer release page figure (nanobody) (developer release page)
Targets 9 · Designs tested — · Hits 6 · Success 6/9 · Best affinity —

Denominator basis: Release page states nine novel targets
Kept only as that version's own figure; it disagrees with the manuscript denominator (10) and must never be merged with it
Executor: Developer's own account

Self-reported · † · verified · source
Known easy targets (five-target set from another model) (manuscript)
Targets 5 · Designs tested — · Hits — · Success nanobody 4/5 (1:1 binders confirmed for 3); miniprotein 3/5 · Best affinity —

Denominator basis: Five easy targets, each present at 100% identity in at least one bound PDB structure
The site's 4/5 holds for nanobodies only
Executor: CRO-executed

Self-reported · † · verified · author-hosted manuscript (link not published)
Partner design campaigns (small molecule / peptide / enzyme / disordered region) (manuscript)
Targets — · Designs tested — · Hits — · Success — · Best affinity —

Denominator basis: The abstract cites eight design campaigns with functional and affinity readouts across several targets; the manuscript does not describe all of them as external partner-lab work
No figures given: the published 'external labs' framing is unsupported
Executor: Partner laboratories, attribution unverified

Self-reported · † · †unverified · author-hosted manuscript (link not published)

Background

  • Organization: The model comes from a US university AI-health clinic together with the Boltz team. Boltz PBC, a public benefit corporation, launched 2026-01-08 in Cambridge, Massachusetts; it pairs open research models with closed enterprise models and does not develop therapeutics itself. No public headcount.
  • Funding: $28M seed led by Amplify, a16z and Zetta Venture Partners with further funds and angels; a $125M valuation appears only in second-hand reports and is not adopted.
  • Products and partnerships: 2026-01-08 multi-year collaboration with a large pharma company (fine-tuning on its historical data, base models stay open, terms undisclosed); 2026-06-18 collaboration with another pharma company granting model and API access, terms undisclosed. The closed successor BoltzProt-1 (2026-06-16) is said to raise hit rates nearly threefold, while a second-hand report says 2.4x; the figures conflict and are unverified.
  • Related models: Boltz-1 / Boltz-1x, Boltz-2, BindCraft, Latent-X / Latent-X2, Chai-2

Reading notes

  1. The same campaign has two denominators (release page nine targets, manuscript ten); read them by version and never merge them into "6/9 in both modalities".
  2. Rates are per target, not per design: 60,000 designs per target, only the top 15 assayed.
  3. Wet-lab work was CRO- and partner-executed and developer-compiled, so it is not independent replication; the "26 external labs" framing has no source.
  4. No site benchmark entry exists, so comparability rests on the wet lab and on the same easy-target set used by Chai-2.

Caveats

  1. Zero pairings: the manuscript reports no site benchmark.
  2. Release page and manuscript disagree on target and hit counts; listed per version.
  3. Partner campaign target counts and success rates are not given as numbers (unverified).
  4. The successor BoltzProt-1 is closed and its claimed gain is reported two conflicting ways.

Sources

  1. author-hosted manuscript (link not published) (preprint)
  2. BoltzGen preprint record (preprint)
  3. Research-institute release page (official site)
  4. GitHub releases (repository)
  5. BoltzProt-1 note reposted by a third-party platform (second-hand) (second-hand) second-hand

Corrections

5 corrections
FieldWasNowBasisFound
wetlab_validation6/9 for both modalitiesManuscript: nanobody 6/10, miniprotein 5/10; the release page separately records 6/9 (nanobody); listed separately by versionauthor-hosted manuscript (link not published)2026-11-16
wetlab_validationEasy targets 4/5Nanobody 4/5, miniprotein 3/5author-hosted manuscript (link not published)2026-11-16
wetlab_validation26 targets validated by 26 external labsNo number given: the manuscript does not say "external labs"; a CRO ran the workauthor-hosted manuscript (link not published)2026-11-16
release_milestones.paper_v1No bioRxiv entrybioRxiv v1 exists (2025-11-24), not published in a journalsource2026-11-16
first_release2025-112025-10-26 (code and weights public)source2026-11-16

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