AI for Bio · Model

Evo 2

Open DNA language model with million-token context for zero-shot variant scoring and sequence generation

Genomics & RNAMixedUnited StatesActive中文

Overview

NameEvo 2 · aliases: Evo 2, Evo2, Evo2-7B, Evo2-40B
Full nameThe paper uses the model name and sizes only
Ambiguous names (not auto-linked)Evo
FamilyNot a family entry. 7B and 40B are sizes of one model, not family members; the predecessor Evo has no record
OrganizationArc Institute, NVIDIA, Stanford University, University of California, Berkeley, University of California, San Francisco
Organization typeMixed
CountryUnited States
Model classGenome language model (sequence generation + variant effects)
DomainGenomics & RNA (also: Molecule generation)
First release† partly verified
Basis: Joint launch by the institute and a chip vendor (preprint the same month)
Note: The launch date comes from coverage rather than a primary page; the code and weights repositories were created earlier and repository creation is not used as first release source
Other milestones
  • First public evidence: † · Launch coverage with weights available source
  • Preprint v1: † · Preprint (month-level evidence only) source
  • Journal online: · Journal online (received 2025-02-18, accepted 2026-01-22) source
  • Code release: · GitHub v0.5.0 source
  • Weights release: † · Weights public at launch source
  • Official announcement: † · Joint launch blog source
Latest update · journal
Basis: Journal publication
Journal online publication; GitHub v0.5.0 came just before (2026-02-28) and later commits (kernel work, 2026-06-19) are minor. No official source for a next-generation model was found as of the as_of date source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Journal publication and a version release within 12 months
OpennessCode: Open
Weights: Open
Weights, training and inference code and the pretraining dataset are all public and integrated into a chip vendor's bio framework; licence text not retrieved this round
Benchmarks reported2 counted (modified protocol 1, retrospective 1, unconfirmed 1)
Wet-lab validation1 recorded campaign
Unverified fields† first_release, release_milestones.paper_v1, release_milestones.weights_release, benchmarks_reported[2], open_status.code_license, open_status.weights_license
As of2026-11-16

What it does

Autoregressively pretrained on DNA with context up to a million nucleotides across bacteria, archaea, eukaryotes and phages; scores variants zero-shot without functional labels and generates new sequences for design. Viruses infecting humans and complex eukaryotes were excluded from training. Unlike AlphaGenome it is self-supervised rather than supervised on functional tracks; like NTv3 it is a long-context genome model, with broader species coverage.

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. 2 of 3 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
11 ClinVar hold-out protocolDeveloper report, modified protocol—Self-reportedIts own subsets, figure-only †partly verified · evidence
37 TraitGymRetrospective dataset use—Third-partyEvaluated on this set by the benchmark's authors. · evidence
39 RNAGymUnconfirmed (not counted)—Unconfirmed sourceThe site matrix records a board entry while this page's figure is quoted via a third-party paper; the board entry awaits re-checking. †disputed · evidence

Other evaluations (outside the 44)

  • Single-gene saturation mutagenesis classification (Own task set) Self-reported: A supervised classifier on the model's representations beats its own zero-shot scores; a developer-defined setup
  • Third-party controlled mitochondrial and nuclear variant study (Comparison against baselines only) Third-party: A third-party preprint says the model calls pathogenic variants benign when only flanking sequence changes, arguing sequence regularity drives predictions; the abstract gives no metric. A comparison figure once cited on the site had no locatable source and is removed
  • Third-party probes on frozen embeddings (Comparison against baselines only) Third-party: A third party trained covariance probes on frozen embeddings and claims parity or better with dedicated predictors; these are supervised probes, not zero-shot, and must not be read with the table above. The competitor list once paraphrased is not in the abstract and is removed
  • BEND (Public benchmark not among the 44) Self-reported: The original paper does not report this benchmark

Developers report: one open all-domains sequence model both scores variants zero-shot and generates designed sequences, with a particular edge on indel-type variants competitors cannot score.

Independent evidence: A third party ran the model under the TraitGym protocol, and two further third-party studies note misclassification when only flanking sequence changes and that high probe scores do not equal zero-shot ability.

Training data

Group in Figure 7: Public data + wet lab for validation only.

ClassValueDetail
a. Public databasesYesA public pan-genome dataset (prokaryotic, eukaryotic and phage sequences); token counts and nucleotide counts are kept as two separate framings per the paper source
b. Private or licensed dataNot disclosed
c. Own wet-lab data used for trainingNo source
d. Wet lab used only for validationYesDesigned sequences integrated into cells and checked by chromatin-accessibility sequencing, for validation only 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.

Generated-DNA accessibility validation (journal paper)
Targets — · Designs tested 36 · Hits 33 · Success 33/36; of 24 designs requiring differential accessibility, 4/24 exceeded twofold and 1/24 threefold · Best affinity —

Denominator basis: 36 constructs (5 in one cell line, 31 in another), per the figure legend
Success is defined by accessibility matching the designed pattern, not a functional phenotype
Executor: New experiments by the developing institute; the method paper is also the wet-lab paper

Self-reported · · verified · source

Background

  • Organization: Led by a California non-profit research institute (founded 2021 with over $650M committed at launch) whose core investigators hold multi-year renewable terms alongside university appointments; no citable official headcount exists, so none is given. Partners are a listed chip company and three US universities.
  • Funding: Non-profit, endowment-funded; 2024 budget figures differ between sources so no single number is given. Compute came from the partner chip company as 2,000 accelerators on a cloud platform (not "over 2,000").
  • Products and partnerships: Builds on the institute's earlier prokaryote and phage genome model; the institute's single-cell line includes STATE and Stack; comparators are AlphaGenome and NTv3; a third-party company runs probing studies on its frozen embeddings.
  • Related models: AlphaGenome, NTv3, STATE, Stack

Reading notes

  1. On ClinVar the edge is on indel-type variants competitors cannot score; it is not first on coding SNVs, so "beats everything" paraphrases do not hold, and values appear only in figures.
  2. High probe scores on embeddings are not zero-shot ability and must be read separately.
  3. Third parties find dedicated methods stronger on mitochondrial and missense variants, showing a trade-off between all-domains coverage and single-task accuracy.
  4. The RNAGym "board" entry is only a figure quoted by another paper and awaits verification.

Caveats

  1. Site benchmark values sit in figures or come from third parties; the text gives no citable numbers.
  2. The RNAGym pairing is unconfirmed and the matrix label disagrees with the benchmark page.
  3. Wet-lab success means the accessibility pattern matched, not a functional phenotype.
  4. The launch date and preprint date rest on coverage and month-level evidence; partly verified.

Sources

  1. Journal paper (journal)
  2. Journal full text (public archive) (journal)
  3. Code repository releases (repository)
  4. Pretraining dataset card (package registry)
  5. Third-party benchmark preprint (third-party paper)
  6. Third-party controlled-study preprint (third-party paper)
  7. Third-party probing preprint (third-party paper)

Corrections

4 corrections
FieldWasNowBasisFound
other_evaluationsTwo comparison numbers from a third-party mitochondrial evaluationRemoved: no source can be found on public pagessource2026-11-16
other_evaluationsProbe study "beats four named models"Removed: the abstract does not list the competitorssource2026-11-16
training_data.a_publicMore than 128,000 genome and metagenome datasetsRemoved: neither the paper text nor the dataset card uses this figuresource2026-11-16
background.fundingMore than 2,000 accelerator cards2,000 accelerator cardssource2026-11-16

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