Overview
| Name | Latent-X / Latent-X2 · aliases: Latent-X, Latent-X2 |
|---|---|
| Full name | Two model generations in one record, so there is no single full name |
| Ambiguous names (not auto-linked) | Latent-Y |
| Family | Covers Latent-X and Latent-X2 only; the Latent-Y design agent built on top is a separate product and its campaign figures belong to the Latent-Y preprint.
|
| Organization | Latent Labs |
| Organization type | AI-native biotech |
| Country | United Kingdom |
| Model class | Design (binder / antibody / macrocycle) |
| Domain | Proteins (also: Antibodies, Molecule generation) |
| First release | † partly verified Basis: Latent-X company press release (the preprint, 2025-07-25, is listed separately) Note: Unverified: that date belongs to the predecessor Latent-X while the X2 press release is 2025-12-16; treated as a family entry source |
| Other milestones | |
| Latest update | † partly verified · product update Basis: Company press release: Latent-Y opened to researchers worldwide Latent-Y, the antibody design agent built on X2 (launched 2026-03-23), opened to researchers worldwide with free daily allowances; no new X2 version or weights were found. Note: A product event, not a model version update to X2 source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Product release events within 12 months, though no model version update |
| Openness | Code: Closed Weights: Closed Weights are closed and access runs through a no-code platform; X2 requires an application or a commercial agreement |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | 3 recorded campaigns |
| Unverified fields | † first_release, latest_update, training_data.a_public, training_data.b_private_licensed, training_data.c_own_wetlab_for_training |
| As of | 2026-11-16 |
What it does
Latent-X targets mini-binders and macrocyclic peptides; X2 extends to drug-like antibodies (VHH, scFv) and reports unoptimised developability measures plus one ex vivo immunogenicity assessment. It shares the closed platform route with Chai-2 and JAM/JAM-2; unlike open models such as BoltzGen, its figures cannot be reproduced externally.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: Latent-X mini-binder / macrocycle campaign; Latent-X2 VHH/scFv campaign; Developability panel; Immunogenicity assay; Latent-Y agent campaigns
Other evaluations (outside the 44)
- Latent-X mini-binder / macrocycle campaign (Wet-lab validation) Self-reported: A self-built target set with 30-100 designs per target, compared against other design models under matched conditions (self-reported)
- Latent-X2 VHH / scFv campaign (Wet-lab validation) Self-reported: Eighteen targets with 4-24 designs per target per modality and binders on nine; macrocycle binders were also made for one hard target
- Developability panel (Own task set) Self-reported: Expression, aggregation, polyreactivity, hydrophobicity and thermal stability versus approved antibodies, with no optimisation or selection
- Immunogenicity assay (Case study) Self-reported: T-cell proliferation and cytokine release on de novo VHH binders for one target in a human donor panel
- Latent-Y agent campaigns (user or partner report) Self-reported: Target-level success across nine nanobody campaigns, sourced to the Latent-Y preprint rather than the X2 paper
Developers report: Developers report zero-shot VHH and scFv designs reaching picomolar to nanomolar affinity from tens of designs per target, with developability close to approved antibodies and low ex vivo immunogenicity (self-reported).
Independent evidence: No independent evidence found. Weights are closed and no third-party evaluation or independent replication was found
Training data
Group in Figure 7: Public data + wet lab for validation only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Not disclosed | Training data sources and scale are not disclosed source |
| b. Private or licensed data | Not disclosed | Whether private or licensed data is used is not disclosed source |
| c. Own wet-lab data used for training | Not disclosed | Not disclosed; must not be written up as 'public data only' source |
| d. Wet lab used only for validation | Yes | Yes: the wet lab is used for validation and no feedback into training is claimed 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: A self-built target set with 30-100 designs per target; the source gives only an aggregate hit rate with no per-target numerators or denominators, so no rate is quoted
Belongs to the predecessor model, and the head-to-head comparisons were run by the developer, not independently
Denominator basis: Eighteen targets with only 4-24 designs tested per target per modality
The source says VHH and scFv and never 'nanobody'; self-reported with no independent replication
Denominator basis: Nine targets with candidates generated by X2 and no human filtering; per-target hit rates run about 1%-28% of tested sequences
Source attribution corrected: these figures come from the Latent-Y preprint, not the 2025-12 press release and not the Latent-X2 paper
Background
- Organization: The company was founded in 2023, headquartered in London with a San Francisco team, and positions itself as an AI-native protein design company; headcount is about 20 on aggregator data (2026-03, profile counts only).
- Funding: It emerged from stealth in 2025-02 with $50M, including a $40M Series A co-led by two funds with several other investors; no later round has been announced.
- Products and partnerships: The products are Latent-X, Latent-X2 and the Latent-Y design agent launched 2026-03 and opened to researchers worldwide in 2026-07; no pharma deal values are public.
- Related models: Chai-2, JAM / JAM-2, AlphaProteo, BoltzGen
Reading notes
- The predecessor's comparison with AlphaProteo is self-reported on company-chosen targets and is not better on every target.
- Binders on 9 of 18 targets is a target-level rate, a different measure from per-target hit rates.
- Neither the training data nor the developability data can be checked externally.
- Both dates hold but mean different things: the company press release is 2025-12-16 and the preprint 2025-12-23.
Caveats
- Zero pairings: no site benchmark score and no competition entry record.
- One record covers two generations; the first-release date belongs to the predecessor Latent-X and is marked unverified.
- The Latent-Y campaign figures belong to a separate preprint and must not be counted as X2 results.
- Only 4-24 designs are tested per target, so target-level success is highly sensitive to the design budget.
Sources
- Latent-X2 preprint (preprint)
- Latent-X preprint (preprint)
- Latent-X2 company press release (press) · †
- Latent-X company press release (press)
- Latent-Y preprint (source of the campaign figures) (preprint)
- Latent-Y availability announcement (press)
Corrections
2 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| wetlab_validation | The nine campaigns and 67% target success were attributed to the 2025-12 press release or the Latent-X2 paper | Attribution corrected to the Latent-Y preprint and written as 6 of 9 targets | source | 2026-11-16 |
| first_release | The first-release date was taken as Latent-X2's own release date | That date belongs to the predecessor Latent-X; the record becomes a family entry with an explicit boundary (unverified) | 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.