Overview
| Name | Boltz-2 · aliases: Boltz-2 |
|---|---|
| Full name | Official materials use Boltz-2 only; no expanded name |
| Ambiguous names (not auto-linked) | Boltz |
| Family | Not a family entry. Covers Boltz-2 only; Boltz-1/1x and BoltzGen have their own records |
| Organization | MIT, Recursion, Boltz PBC |
| Organization type | Mixed · Jointly released by a university lab and a listed biotech company, now maintained by a public benefit corporation |
| Country | United States |
| Model class | Co-folding plus affinity prediction |
| Domain | Structure & co-folding (also: Small molecules) |
| First release | · code release Basis: GitHub release v2.0.0; preprint 2025-06-18 source |
| Other milestones | |
| Latest update | † unverified · product update Basis: Product-line event: two new models and a hosted API went live (secondary sources only) The developers announced two new models and a hosted API running this model; the open repository's last release was 2025-09-08 and the weights themselves are unchanged (unverified) Note: The date appears only in secondary coverage; the developers' own page was not retrieved (unverified) source The most recent sourced event, not a claim that nothing changed since. |
| Status | Maintained only · No open-repository release since 2025-09; the 2026 product event is unverified |
| Openness | Code: Open · MIT Weights: Open · MIT Weights and inference code are MIT-licensed; as of 2026-11-16 the repository still marks the training code and evaluation scripts as coming soon |
| Benchmarks reported | 3 counted (board 2, retrospective 1, cited only 1, unconfirmed 2) |
| Wet-lab validation | 1 recorded campaign |
| Unverified fields | † latest_update, training_data.b_private_licensed, release_milestones.weights_release, benchmarks_reported[1], benchmarks_reported[2], benchmarks_reported[3], benchmarks_reported[5], background.products_partnerships |
| As of | 2026-11-16 |
What it does
On the structure side it adds method conditioning and more control over the predecessor; on the affinity side a head on top of the structure module outputs binder or non-binder plus binding strength for hit finding and hit-to-lead ranking. That affinity side is what separates it from Boltz-1/1x and Chai-1, and Nesso-1, TerraBind and AQAffinity reuse its public-data recipe.
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. 3 of 6 pairings are counted.
| Benchmark | Relation | Eval set / protocol | Source type | Note and evidence |
|---|---|---|---|---|
| 31 FoldBench | Official leaderboard or results | maintainer run on post-2024-01 subset | Board | Maintainer ran Boltz-2 on post-2024-01 subset (training cutoff later; reference only) †partly verified · evidence |
| 34 AbBiBench | Official leaderboard or results | registered interface-score entry | Board | Benchmark-registered ipSAE entry; not structure accuracy †unverified · no source link |
| 01 CASP | Retrospective dataset use | affinity track, retrospective (2 targets / 140 pairs) | Third-party | Retrospective CASP16 affinity data (2 targets, 140 pairs); not a blind participation · evidence |
| 30 PLINDER / PINDER / Runs N' Poses | Cited as comparison only (not counted) | Runs N' Poses | Third-party | Quoted in the Pearl paper and flagged not directly comparable; its official evaluation docs omit this benchmark †unverified · no source link |
| 12 DUD-E / LIT-PCBA | Unconfirmed (not counted) | decoy-set screening in a third-party paper | Unconfirmed source | Third-party paper uses DEKOIS 2.0 etc., not clearly DUD-E/LIT-PCBA; second-hand †disputed · evidence |
| 27 Polaris / ASAP / OpenADMET | Unconfirmed (not counted) | — | Unconfirmed source | Site page says it matched the antiviral top 5 without fine-tuning; no numbers; participation not openly verified †disputed · no source link |
Other evaluations (outside the 44)
- Free-energy calculation comparison subset (Own task set) Self-reported: Correlation against a physics-based method on a four-target subset, self-reported; scores are not shown here
- Hit-to-lead and screening sets, self-reported (Own task set) Self-reported: Self-reported results on a public hit-to-lead dataset and a high-throughput screening dataset, neither a site benchmark
- Third-party prospective evaluation (third party prospective evaluation) Third-party: A prospective study on about 557 co-crystal complexes finds its affinity score best for early enrichment while its poses trail the contemporary closed model
- Third-party protein-protein affinity study (third party evaluation) Third-party: A paper finds it still below a sequence baseline after fine-tuning
Developers report: near physics-based free-energy accuracy on a four-target subset at about three orders of magnitude lower cost, matching the top entries of a public challenge without fine-tuning.
Independent evidence: Third-party prospective work finds the affinity score useful for enrichment but insufficient alone for virtual screening; other work argues screening results are inflated by training overlap and that it trails a sequence baseline on protein-protein tasks.
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | PDB (to 2023-06-01), MD ensembles, distillation plus public affinity libraries such as ChEMBL; totals differ by source source |
| b. Private or licensed data | Not disclosed | None in the public version; from 2026-01 a pharma partner's historical data is used for refinement, with no statement on the public version |
| c. Own wet-lab data used for training | No | |
| d. Wet lab used only for validation | No | No new wet lab of its own; a partner's eight internal assays serve as a held-out blind test and one prospective item is a simulation 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.
Denominator basis: Eight internal partner assays, using existing experimental data as a held-out blind test rather than wet lab run for this model
Partner-lab executed and developer-compiled, so not an independent replication; the correlation threshold is a metric and is not shown here
Background
- Organization: At release it was co-authored by two research units of a US university and a listed biotech company that supplied data and compute; it is now maintained by a public benefit corporation spun out of that university, which says it does not develop its own drugs.
- Funding: The maintainer launched publicly on 2026-01-08 with a $28M seed round co-led by three firms; valuation undisclosed.
- Products and partnerships: A platform product launched the same day, alongside a pharma collaboration building custom structure, affinity and biologics-design models on that company's data (custom development, per second-hand analysis); the two new models and hosted API of 2026-06-16 appear only in secondary coverage (unverified).
- Related models: Boltz-1 / Boltz-1x, BoltzGen, Nesso-1, TerraBind, AQAffinity, IsoDDE
Reading notes
- The free-energy comparison holds only on a four-target subset, and only high-sequence-identity targets were removed; overlap with other evaluation sets is not reported set by set.
- Only a minority of the eight partner blind assays clear the threshold, so performance on new targets is unstable and the subset result cannot be extrapolated.
- Because of its later cutoff the structure side is often excluded from PoseBusters-style comparisons, so no claim either way about structure quality follows.
- The 2026 product event comes from secondary coverage, and the weights themselves are unchanged since 2025-09.
Caveats
- Training code and evaluation scripts are still marked coming soon in the repository, so full reproduction of the paper is not yet possible.
- The most recent update event has secondary sources only and must display as unverified.
- Three pairings rest only on the site's own pages (unverified).
- One pairing uses a different decoy set and its membership of that family is still undecided.
- The wet-lab item is a held-out blind test on a partner's existing data, neither new experiments nor an independent replication.
Sources
- The model's preprint (preprint)
- Paper full text (public archive) (paper)
- Official repository (repository)
- Official repository releases (repository)
- Institute release note (official site)
- Third-party prospective evaluation preprint (preprint)
- Third-party re-run paper (third-party paper)
- Coverage of the 2026 product event (second-hand) (second-hand) second-hand
Corrections
3 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| latest_update | The 2026-06-16 product event was presented as settled fact | Marked unverified: secondary coverage only, with the developers' own page not retrieved | source | 2026-11-16 |
| other_evaluations | A free-energy correlation value relayed by a second-hand page | Corrected to the paper's own figure, and this record shows no concrete scores | source | 2026-11-16 |
| wetlab_validation | The partner assays were treated as this model's own wet-lab validation | Reclassified as a held-out blind test on existing data, partner-lab executed, with one prospective item a simulation rather than an experiment | source | 2026-11-16 |
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.