Overview
| Name | Chai-2 · aliases: Chai-2, Chai-2f |
|---|---|
| Full name | Official materials use Chai-2; the built-in folding module is called Chai-2f |
| Ambiguous names (not auto-linked) | Chai |
| Family | Not a family entry. Covers Chai-2 and its folding module only; Chai-1 and Chai-3 have their own records |
| Organization | Chai Discovery |
| Organization type | AI-native biotech |
| Country | United States |
| Model class | Design (antibody / nanobody / miniprotein) plus built-in folding |
| Domain | Structure & co-folding (also: Antibodies, Molecule generation) |
| First release | · product update Basis: Official research post and press release the same day source |
| Other milestones |
|
| Latest update | † partly verified · successor Basis: The successor was disclosed through a licence announcement Partner-only successor Chai-3 was first disclosed; an earlier post on 2025-11-19 covered full-length antibody design Note: The official news page has no dedicated post for the successor (partially verified) source The most recent sourced event, not a claim that nothing changed since. |
| Status | Product only, no public artifact · Closed and partner-only with no public weights; the successor is partner-only too |
| Openness | Code: Closed Weights: Closed Closed and partner-only, with no public weights or inference code |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | 3 recorded campaigns |
| Unverified fields | † wetlab_validation[2], release_milestones.paper_v1, latest_update, training_data.a_public, background.products_partnerships |
| As of | 2026-11-16 |
What it does
Given a target and epitope residues it generates antibodies, nanobodies or miniproteins zero-shot without a known starting binder, with a folding module included for structural assessment. Its role differs from general co-folding in Chai-1, and BindCraft, BoltzGen and RFdiffusion2/3 are open counterparts whose hit-rate definitions differ.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: Antibody/nanobody wet-lab campaign; Miniprotein wet-lab test; Antibody-antigen folding (Chai-2f); Developability / polyreactivity; Full-length IgG update (2025-11)
Other evaluations (outside the 44)
- Antibody and nanobody wet-lab campaign (Wet-lab validation) Self-reported: 52 novel antigens with up to 20 designs per target per format, hits by binding signal, compared with earlier de novo antibody methods
- Miniprotein wet-lab test (Wet-lab validation) Self-reported: Five targets with 20-25 designs each, compared with the published results of two earlier methods
- Antibody-antigen folding module self-evaluation (Comparison against baselines only) Self-reported: On 152 low-homology antibody-antigen complexes (255 interfaces) against the predecessor and two baselines; scores are not shown here
- Developability and polyreactivity own-task set (Own task set) Self-reported: Polyreactivity assays on the designs plus developability and immunogenicity scores against approved antibodies and paired natural sequences
- 2025-11 full-length antibody update (user or partner report) Self-reported: A developer update on full-length antibody developability, known only through secondary reports; no figures given (unverified)
Developers report: zero-shot double-digit per-design hit rates from small design sets, binders for about half of the novel antigens, and antibody-antigen docking accuracy better than the predecessor.
Independent evidence: A competitor report (JAM/JAM-2) re-tested 16 overlapping targets and compares bind rates; no neutral third-party replication was found.
Training data
Group in Figure 7: Public data + wet lab for validation only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Not disclosed | |
| b. Private or licensed data | Not disclosed | |
| c. Own wet-lab data used for training | Not disclosed | The technical report does not say whether it has own training data |
| d. Wet lab used only for validation | Yes | Design validation on 52 targets; no public statement that results were fed back into training 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: 52 novel antigens with up to 20 designs per target per format (VHH and scFv), hits called by a binding-signal threshold
Self-reported on self-chosen targets with no independent replication; the per-design and per-target rates must not be mixed
Denominator basis: Five targets with 20-25 designs each, read out by binding signal
Comparisons with other methods' published figures are citations, not re-runs; not an independent replication
Denominator basis: Secondary reports only with no primary text located, so neither numerator nor denominator is given (unverified)
The two ratios published earlier could not be matched to any primary source, so no figures are given under this round's rules (unverified)
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: About $30M seed in 2024-09, a $70M Series A on 2025-08-06, a $130M Series B at $1.3B on 2025-12-15 and a $400M Series C at $3.8B in 2026-07; the company gives no official cumulative figure.
- Products and partnerships: The line runs from open-weights Chai-1 through this design model to partner-only Chai-3; four pharma collaborations through 2026 are press releases with no verifiable performance data and undisclosed terms.
- Related models: Chai-1, Chai-3, BindCraft, BoltzGen, Latent-X / Latent-X2, JAM / JAM-2
Reading notes
- Every figure is a developer self-report on self-chosen targets, with no third-party replication and no result in a common public challenge.
- Its hit-rate definitions (per design or per target, binding or function) differ from other design models, so cross-model ranking has no basis.
- It is closed and discloses no training data, so overlap between training and test targets cannot be assessed.
- The collaboration announcements are press releases covering access only, and the successor has no public benchmark at all.
Caveats
- Zero pairings: this record has no site benchmark entry, so comparability rests on self-reported wet lab.
- The 2025-11 full-length antibody update is treated as unverified and carries no figures.
- Only a DOI-string date exists for the preprint v1, which this round excludes, so that milestone is left empty.
- Second-hand paraphrases of the successor doubling success rates are inconsistent and are not adopted here.
- The wet lab is developer-reported with formats and targets of its own choosing, not an independent evaluation.
Sources
- Official research post (release page) (official site)
- Companion technical report preprint (preprint)
- Official news list (official site)
- Most recent pharma collaboration press release (press)
- Series B coverage (second-hand) (second-hand) second-hand
Corrections
3 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| wetlab_validation | The full-length antibody update gave two concrete ratios | No figures given: they could not be matched to any primary source (unverified) | source | 2026-11-16 |
| other_evaluations | The folding self-evaluation and the design hit rates were read together | Separated by category: wet-lab validation, comparison versus baseline only, own task set and partner report each stand alone | source | 2026-11-16 |
| background.products_partnerships | Paraphrased performance claims for the successor were treated as fact | Changed to: the collaboration announcements are press releases with no verifiable performance data | 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.