AI for Bio · Model

Tahoe-x1

Single-cell foundation model: masked-pretraining encoder trained at scale including perturbation data

Single-cell & perturbationAI-native biotechUnited StatesMaintained only中文

Overview

NameTahoe-x1 · aliases: Tahoe-x1, Tx1-70M, Tx1-1.3B, Tx1-3B
Full nameThe preprint and model card use the product name only
Ambiguous names (not auto-linked)Tahoe
FamilyNot a family entry. The three parameter sizes are configurations of one model, not family members
OrganizationTahoe Therapeutics
Organization typeAI-native biotech
CountryUnited States
Model classSingle-cell foundation model (masked-pretraining encoder)
DomainSingle-cell & perturbation
First release · code release
Basis: First code and weights release (v0.1.0); the preprint followed on 2025-10-23
Note: The model-hub repository was created 2025-08-25 with a README only; repository creation is not used as first release source
Other milestones
  • First public evidence: · v0.1.0 release source
  • Preprint v1: · Preprint v1 (10.1101/2025.10.23.683759) source
  • Code release: · GitHub v0.1.0 source
  • Weights release: † · Weights released with the code under a permissive licence source
Latest update · maintenance only
Basis: Perturbation tutorial added to the repository
A perturbation tutorial was added to the repository, a maintenance change; no new version or journal publication was found this round. This is the latest sourced event source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · Maintenance commits only, with no new version or journal publication
OpennessCode: Open · Apache-2.0
Weights: Open · Apache-2.0
All three weight sizes and the code are public under a permissive licence
Benchmarks reported1 counted (modified protocol 1)
Wet-lab validation1 recorded campaign
Unverified fields† release_milestones.weights_release, other_evaluations[2], background.products_partnerships
As of2026-11-16

What it does

Masked language modelling with an encoder, with perturbation data deliberately included in pretraining; downstream tasks span gene essentiality, cell-type classification and perturbation response. The developers claim several-fold to tens-of-fold better compute efficiency than existing cell-state models - a company statement, not independently verified. It shares a perturbation corpus with STATE but is a representation-learning foundation model rather than a response predictor, and differs again from the in-context route of Stack.

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
24 Virtual Cell Challenge and perturbation benchmarksDeveloper report, modified protocol—Self-reportedAn own benchmark, explicitly not VCC. · evidence

Other evaluations (outside the 44)

  • Public gene-essentiality dataset (Public benchmark not among the 44) Self-reported: Global and context-specific essentiality prediction on public CRISPR screening data, which is a different task from perturbation response; the abstract and model card state conclusions without numbers
  • Cancer hallmark genes and cell-type classification (Own task set) Self-reported: The abstract claims state-of-the-art on all four tasks with no citable figures
  • Compute-efficiency claim (Comparison against baselines only) Self-reported: The company claims several-fold to tens-of-fold higher efficiency, not independently verified

Developers report: including their own perturbation data in pretraining puts the model at state of the art on essentiality, hallmark genes, cell typing and held-out-context perturbation response at once, with better compute efficiency.

Independent evidence: No independent evidence found. No third-party evaluation or replication was found; another institute's model uses the same data producer's corpus, but that is not an evaluation of this model

Training data

Group in Figure 7: Own wet-lab training data.

ClassValueDetail
a. Public databasesYesTwo public single-cell atlases and count collections, totalling 266 million single-cell profiles with the in-house corpus source
b. Private or licensed dataNot disclosed
c. Own wet-lab data used for trainingYesIn-house perturbation corpus: about 100 million profiles, 1,100 small-molecule perturbations, 50 cancer cell lines; the data producer and the model developer are the same company source
d. Wet lab used only for validationNo 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.

In-house perturbation corpus scale (training data, not a validation campaign) (primary paper v3)
Targets — · Designs tested — · Hits — · Success — · Best affinity —

Denominator basis: About 100 million profiles, 1,100 small-molecule perturbations, 50 cancer cell lines; 266 million cells in pretraining
The perturbation count is 1,100 this round; the ~1,200 figure circulating in press releases is not used, and the same fix applies to the training-data line of STATE
Executor: Produced on the developer's own experimental platform, used for training rather than validation

Self-reported · · verified · source

Background

  • Organization: A single-cell data and model company founded in 2022 on technology from a US university, headquartered in California; sources date its renaming to either 2024-04 or 2025-04. Its core assets are its experimental platform and in-house perturbation dataset. No citable headcount, so none is given.
  • Funding: A $12M seed in 2022-12 and a $30M Series A in 2025-08 (one fund leading with six further investors), $42M in total, with the round earmarked for a billion-scale single-cell dataset.
  • Products and partnerships: Sample prep and sequencing for the in-house corpus were done by two instrument and reagent companies, and the corpus was open-sourced as the first dataset in a research institute's cell atlas; in 2026-01 the company, that institute and a biohub announced a larger joint perturbation dataset (press-release framing: over 120 million single-cell data points, about 50 cell lines, 1,400 chemical scaffolds at three doses each and 100 cytokine perturbations, shared among the three parties before open release); also in 2026-01 it announced a joint venture with an antibody platform company for antibody-drug conjugates, starting with two programmes and undisclosed terms. On-site relation: STATE trains on the same corpus.
  • Related models: STATE, Stack, X-Cell

Reading notes

  1. The essentiality evaluation uses public CRISPR data, a different task from perturbation response, and the four state-of-the-art claims come without numbers.
  2. Producing its own data removes third-party dependence but brings the risk that training and evaluation share one distribution.
  3. The efficiency multiple is a company statement that has not been independently verified.
  4. With no blind challenge score it cannot be compared with other perturbation models on the same blind test.

Caveats

  1. The only site pairing is self-reported under a modified protocol that the developers say is not the challenge.
  2. No journal version; the latest event is repository maintenance only.
  3. Perturbation counts conflict between press releases and the paper; the paper's figure is used.
  4. In-house data enters the training set, so the within-distribution risk must be stated when reading results.

Sources

  1. Preprint (preprint)
  2. Weights and model card (package registry)
  3. Code repository releases (repository)
  4. In-house perturbation corpus paper (preprint)
  5. Corpus open-sourcing announcement (third-party institute) (official site)

Corrections

2 corrections
FieldWasNowBasisFound
training_data.c_own_wetlab_for_trainingAbout 1,100–1,200 drugs1,100 small-molecule perturbations (about 1,200 appears only in press releases)source2026-11-16
first_release2025-10 (preprint)2025-10-10 (first code and weights version)source2026-11-16

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