AI for Bio · Model

Chai-1

Open-weights co-folding model that stays stable without MSAs and accepts experimental restraints

Structure & co-foldingAI-native biotechUnited StatesSuperseded by a successor中文

Overview

NameChai-1 · aliases: Chai-1, chai-lab
Full nameOfficial materials use Chai-1 only; no expanded name
Ambiguous names (not auto-linked)Chai
FamilyNot a family entry. Covers Chai-1 only; Chai-2 and Chai-3 have their own records
OrganizationChai Discovery
Organization typeAI-native biotech
CountryUnited States
Model classCo-folding (protein / nucleic acid / small molecule / glycan)
DomainStructure & co-folding
First release · product update
Basis: Official release post with the repository and web server at the same time; preprint 2024-10-11 source
Other milestones
  • Preprint v1: · Preprint v1 source
  • Code release: † · Repository made public source
  • Weights release: † · Weights provided at launch source
  • Official announcement: · Official release post source
Latest update† partly verified · successor
Basis: The successor was first disclosed through a licence announcement
Chai-3 was first disclosed as a partner-only successor with no public weights; the open repository's last release was 2025-03-18
Note: Outlets report dates between 06-02 and 06-05 for this event (partially verified) source
The most recent sourced event, not a claim that nothing changed since.
StatusSuperseded by a successor · The company moved to successor design models; no open release since 2025-03
OpennessCode: Open · Apache-2.0 (from v0.4.0, 2024-11-27)
Weights: Open · Apache-2.0 (from v0.4.0, 2024-11-27)
Licence change: † partly verified · code release
Basis: From chai-lab v0.4.0 both code and weights are Apache-2.0 source
Code and weights share one licence allowing commercial use; the initial terms before v0.4.0 were not re-verified this round
Benchmarks reported5 counted (board 1, self 1, retrospective 3, cited only 1)
Wet-lab validationNone recorded
Unverified fields† benchmarks_reported[0], benchmarks_reported[2], latest_update, release_milestones.weights_release, training_data.b_private_licensed
As of2026-11-16

What it does

It predicts protein, nucleic-acid, small-molecule and glycan complexes. Its distinctive mode is single-sequence inference: accuracy drops only modestly without MSAs, which suits novel targets and antibodies, and experimental restraints can be supplied. It is one open-weights alternative to AlphaFold 3, and splits prediction from design with its successor Chai-2.

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. 5 of 6 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
31 FoldBenchOfficial leaderboard or resultsmaintainer leaderboardBoardMaintainer leaderboard · evidence
05 PoseBustersDeveloper report, official protocolPoseBusters V1 (428), official setSelf-reportedV1 self-report (preprint, quoted via reviews); the V2 numbers are third-party · evidence
01 CASPRetrospective dataset useprotein monomers of one past seasonThird-partyCASP15 monomer LDDT, second-hand quote only; post hoc †unverified · no source link
33 PoseXRetrospective dataset usebenchmark authors' own setsThird-partySame table as above. †partly verified · evidence
35 FLAb / FLAb2Retrospective dataset usezero-shot thermostability eval in the benchmark's own paperThird-partyThe FLAb2 paper evaluates this model zero-shot on thermostability; two site surfaces label the source type differently. †partly verified · evidence
30 PLINDER / PINDER / Runs N' PosesCited as comparison only (not counted)Runs N' Poses, best@5Third-partyQuoted from the Pearl paper †unverified · no source link

Other evaluations (outside the 44)

  • Third-party MSA-free robustness evaluation (third party evaluation) Third-party: A third-party evaluation concludes it is the only method that stays stable without MSAs
  • Third-party tables outside the maintainer board (third party evaluation) Third-party: Several successor-model papers include it in comparison tables, all third-party re-runs under differing protocols
  • Self-reported results quoted via reviews (Comparison against baselines only) Self-reported: The preprint's self-reported results were read only through reviews this round, not from the full text

Developers report: co-folding accuracy on par with the contemporary closed model, little loss without MSAs, and conditioning on experimental restraints.

Independent evidence: Third-party re-runs under stricter protocols place it below its self-reported level; one third-party evaluation independently confirms its MSA-free robustness, and a maintainer leaderboard includes it.

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesPDB and other public sources; no cutoff date or entry counts given officially source
b. Private or licensed dataNot disclosed
c. Own wet-lab data used for trainingNot disclosed
d. Wet lab used only for validationNoNo wet-lab work reported

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The company was founded in 2024-03 in San Francisco, worked early on from an AI lab's offices, and positions itself as an AI-native molecular design platform selling models and custom models rather than running its own pipeline; headcount undisclosed.
  • Funding: A $30M seed in 2024-09, a $70M Series A in 2025-08, a $130M Series B at $1.3B on 2025-12-15 and a $400M Series C at $3.8B in 2026-07.
  • Products and partnerships: The product line runs from this open-weights model to Chai-2 for design and partner-only Chai-3; collaborations with four large pharma companies include custom models and design kits, all with undisclosed terms.
  • Related models: Chai-2, Chai-3, AlphaFold 3, Boltz-1 / Boltz-1x, Boltz-2, Protenix

Reading notes

  1. The self-report uses the looser official protocol while third-party re-runs under stricter rules land much lower, so the self-report should not be read as beating the contemporary closed model.
  2. Its MSA-free robustness comes from a third-party evaluation, which is firmer than the self-report and is its most distinctive finding.
  3. Training-data information is minimal, with no cutoff or entry counts, so leakage risk cannot be assessed.
  4. The antibody-property pairing carries inconsistent source labels in two places on the site; reading it as the benchmark side's own paper is safer.

Caveats

  1. The self-reported results were read only via reviews, not the original full text (partially verified).
  2. Two pairings rest only on the site's own pages (unverified).
  3. Outlets differ by a few days on the successor's disclosure date (partially verified).
  4. The initial weights terms were not re-verified; only the Apache-2.0 status from v0.4.0 is confirmed.

Sources

  1. Official release post (official site)
  2. The model's preprint (preprint)
  3. Official repository (repository)
  4. Third-party re-run paper (third-party paper)
  5. Re-run paper by the benchmark's authors (third-party paper)
  6. Paper from the antibody-property benchmark side (third-party paper)
  7. Maintainer leaderboard repository (leaderboard)
  8. Official news list (successor disclosure) (official site)

Corrections

3 corrections
FieldWasNowBasisFound
first_releaseThe first release was given as a 2024-09/10 rangeChanged to the day of the official release post, with the preprint listed as a separate milestonesource2026-11-16
open_statusThe licence was described as Apache-2.0 from launchChanged to Apache-2.0 from v0.4.0 (2024-11-27), with the initial terms not re-verifiedsource2026-11-16
benchmarks_reportedThe antibody-property row carried both third-party and self-report labelsUnified as retrospective dataset use in the benchmark side's own paper, with the label mismatch recordedsource2026-11-16

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