AI for Bio · Model

SeedFold

ByteDance Seed AF3-style co-folding model that scales by widening the Pairformer, with a linear triangle-attention variant

Structure & co-foldingBig-tech labChinaProduct only, no public artifact中文

Overview

NameSeedFold · aliases: SeedFold, SeedFold-Linear
Full nameNo expanded name
Ambiguous names (not auto-linked)None
FamilyNot a family entry. One model from one paper; SeedFold-Linear is a variant in the same paper, not a separate entry
OrganizationByteDance Seed
Organization typeBig-tech lab
CountryChina
Model classCo-folding (structure prediction)
DomainStructure & co-folding
First release · preprint
Basis: arXiv 2512.24354 v1 submission source
Other milestones
  • First public evidence: · arXiv v1 source
  • Preprint v1: · arXiv v1 source
Latest update · no later event
Basis: arXiv v1; project page and web search show no later event
No later sourced event found (as of 2026-11-16); the paper links a web server, with no code, weights page or changelog found source
The most recent sourced event, not a claim that nothing changed since.
StatusProduct only, no public artifact · Paper and public web server only; no code or weights
OpennessCode: Unclear
Weights: Unclear
The paper states no code or weight licence; a public web server exists, no code or weights page found
Benchmarks reported1 counted (self 1)
Wet-lab validationNone recorded
Unverified fields† open_status.code, open_status.weights
As of2026-11-16

What it does

Keeps the AlphaFold 3 Pairformer-plus-diffusion design but argues performance is limited by hidden width, so it widens the Pairformer; SeedFold-Linear swaps triangle attention for a linear version to cut cost. Where the same company's Protenix reproduces AF3 through data and recipes, SeedFold contributes architecture scaling; SeedProteo later reuses its folding architecture for design.

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
31 FoldBenchDeveloper report, official protocolFoldBench categories (protein-ligand, antibody-antigen, protein-protein, RNA)Self-reportedSelf-reported on FoldBench; arXiv 2512.24354 · evidence

Other evaluations (outside the 44)

  • Width and linear-attention ablations (Own task set) Self-reported: In-paper comparison of the two variants and scaling choices

Developers report: widening the Pairformer beats stacking layers, and linear triangle attention cuts compute while keeping accuracy.

Independent evidence: No independent evidence found. No third-party replication or board-run result found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesPDB 180,540, AFDB 3,326,991 and MGnify distillation 23,075,211 entries, cutoff 2021-09-30 source
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: ByteDance Seed is ByteDance's foundation-model research division; its AI for Science work covers biological foundation models, quantum chemistry and molecular dynamics. A second-hand report (2026-06) says an AI drug-discovery unit absorbing the protein-structure team is being spun out under ByteDance control; whether SeedFold moves is not stated.
  • Funding: Funded from ByteDance's internal budget; no separate fundraising.
  • Products and partnerships: The same company open-sources the Protenix series and PXDesign; SeedProteo reuses the SeedFold architecture. No pharma partnerships disclosed.
  • Related models: Protenix, SeedProteo, AlphaFold 3, Boltz-2

Reading notes

  1. The two variants trade wins across complex categories; on RNA the paper's own table still favours AF3.
  2. All numbers are self-reported on FoldBench without third-party replication and are not comparable with PoseBusters-type figures.
  3. Only public data is used, so gains come from architecture and scale rather than private data.

Caveats

  1. Weight availability is not stated; practical access is unverified.
  2. Same company as Protenix but a different team; their relationship is not described officially.
  3. Organizational spin-out information comes from second-hand reports only.

Sources

  1. SeedFold arXiv 2512.24354 (preprint)
  2. SeedFold project page (official site)
  3. SeedProteo arXiv 2512.24192 (preprint)
  4. Seed AI for Science page (official site)
  5. Report on ByteDance spinning out its drug-discovery unit (second-hand) (second-hand) second-hand

Corrections

1 correction
FieldWasNowBasisFound
benchmarks_reportedNo direct site benchmark pairing; FoldBench outside scopeFoldBench is now a site entry; counted as self-reported under the official protocolsource2026-11-16

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