AI for Bio · Model

Boltz-1 / Boltz-1x

First fully open co-folding model approaching AF3, with 1x adding physical-validity steering

Structure & co-foldingMixedUnited StatesSuperseded by a successor中文

Overview

NameBoltz-1 / Boltz-1x · aliases: Boltz-1, Boltz-1x
Full nameBoth members use short names only; no expanded name
Ambiguous names (not auto-linked)Boltz
FamilyThis record covers Boltz-1 and Boltz-1x only; Boltz-2 and BoltzGen have their own records and are not merged in.
  • Boltz-1 ·
  • Boltz-1x ·
OrganizationMIT, Boltz PBC
Organization typeMixed · Started in a university lab, now maintained by a public benefit corporation
CountryUnited States
Model classCo-folding (protein / nucleic acid / small molecule)
DomainStructure & co-folding
First release · code release
Basis: Boltz-1's first GitHub release v0.1.0; preprint 2024-11-20 source
Other milestones
  • Preprint v1: · Preprint v1 source
  • Code release: · GitHub release v0.1.0 source
  • Weights release: † · Weights published with the first release under MIT source
Latest update · successor
Basis: Successor in the same product line released
Boltz-2 shipped as the same-line successor (adding affinity prediction); Boltz-1x's steering was v1.0.0 on 2025-04-26 source
The most recent sourced event, not a claim that nothing changed since.
StatusSuperseded by a successor · Superseded by its successor but still a common open control
OpennessCode: Open · MIT
Weights: Open · MIT
Code and weights are both MIT-licensed with commercial use allowed; the technical report says the dataset and evaluation sets are released under MIT too
Benchmarks reported2 counted (board 1, retrospective 1, cited only 2)
Wet-lab validationNone recorded
Unverified fields† benchmarks_reported[1], release_milestones.weights_release, background.products_partnerships
As of2026-11-16

What it does

The first fully open co-folding model to approach AlphaFold 3 on public evaluations, covering protein, nucleic-acid and small-molecule complexes; 1x adds steering potentials at inference to cut stereochemical and steric clashes. It is a standard open control in papers such as Chai-1 and Protenix.

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 4 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
31 FoldBenchOfficial leaderboard or resultsmaintainer leaderboardBoardThe maintainer board lists Boltz-1 · evidence
33 PoseXRetrospective dataset usebenchmark authors' own setsThird-partyThe benchmark paper's authors evaluated this model on their own sets. †partly verified · evidence
05 PoseBustersCited as comparison only (not counted)PoseBusters V2, best@5Third-partyNumbers come from another paper, not a self-report †partly verified · evidence
30 PLINDER / PINDER / Runs N' PosesCited as comparison only (not counted)Runs N' Poses, best@5Third-partyNumbers from the Pearl paper †unverified · no source link

Other evaluations (outside the 44)

  • Inclusion in third-party evaluation suites (third party evaluation) Third-party: Two third-party evaluation efforts (a 2026 journal paper and a conference paper) include it in multi-method comparisons, relayed second-hand
  • Qualitative statements in the technical report (Comparison against baselines only) Self-reported: The abstract and README give only qualitative statements of parity with contemporary closed models, with no evaluation values
  • Qualitative finding from a third-party reproduction (third party evaluation) Third-party: A reproduction report says it works for most orthosteric ligands, without summary metrics

Developers report: the first fully open model to match contemporary closed co-folding quality on public evaluations, with code, weights and dataset under a commercially usable licence.

Independent evidence: Every figure on its page comes from third-party re-runs, plus inclusion in a maintainer leaderboard; the developers report no scores themselves, so reproducibility is the firmest evidence.

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesPDB plus distillation; no per-database sizes, dataset released under MIT source
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNoNeither member reports wet-lab work

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The model comes from a US university lab and is now maintained by a public benefit corporation spun out of it; the company launched publicly on 2026-01-08, says it does not develop its own drugs and charges for software and platform access, with headcount undisclosed.
  • Funding: 2026-01 a $28M seed round co-led by three firms with two further funds and angels participating; valuation not officially disclosed.
  • Products and partnerships: Two lines, open models plus closed enterprise models; a 2026-01 multi-year pharma collaboration refines custom models on that company's historical data while the base models stay open; on 2026-06-16 two further proprietary models and a hosted API were announced (secondary coverage only, unverified).
  • Related models: Boltz-2, BoltzGen, AlphaFold 3, Chai-1, Protenix, OpenFold3

Reading notes

  1. The gap between the two members comes mainly from physical-validity steering, so the validity criterion matters a lot and comparisons must name which member is meant.
  2. All concrete figures are third-party re-runs under differing protocols and cannot be read as a strict ranking.
  3. Its value lies in reproducibility and fine-tunability rather than leaderboard position.
  4. The successor's release is a same-line event and does not mean these two members were themselves updated.

Caveats

  1. Family boundary: Boltz-1 and Boltz-1x only; the successor and the design model stand alone.
  2. The developers self-report no benchmark values; all figures are third-party.
  3. One pairing rests only on the site's own page and displays as unverified.
  4. The weights-release date rests on the README with no separate release page (partially verified).
  5. The two proprietary models of 2026-06-16 have secondary coverage only and are not adopted as facts here.

Sources

  1. Official repository (repository)
  2. Boltz-1 preprint (preprint)
  3. Official repository releases (repository)
  4. Third-party re-run paper (third-party paper)
  5. Re-run paper by the benchmark's authors (third-party paper)
  6. Coverage of the maintaining public benefit corporation's launch (second-hand) (second-hand) second-hand

Corrections

3 corrections
FieldWasNowBasisFound
familyThe two members were written up as a single 2024-11 releaseThe family boundary is stated: Boltz-1 on 2024-11-17 and the steering version Boltz-1x on 2025-04-26source2026-11-16
first_releaseFirst release taken from the preprint dateChanged to the earlier first code release, with the preprint kept as a separate milestonesource2026-11-16
benchmarks_reportedThird-party re-run figures were displayed as developer self-reportsTwo rows become comparison-only citations and are not counted as established relationsno source link2026-11-16

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