Overview
| Name | NTv3 · aliases: NTv3, Nucleotide Transformer v3 |
|---|---|
| Full name | Nucleotide Transformer v3 |
| Ambiguous names (not auto-linked) | Nucleotide Transformer, NT |
| Family | Not a family entry. 8M, 100M and 650M are sizes of one generation; the two earlier generations have no entity records |
| Organization | InstaDeep (subsidiary of BioNTech) |
| Organization type | AI-native biotech |
| Country | United Kingdom |
| Model class | Genome language model / sequence-to-function model |
| Domain | Genomics & RNA (also: Molecule generation) |
| First release | · paper revision Basis: Preprint v1 posted date (version API); a technical report PDF appeared the same month Note: The earlier 2025-12-22 was inferred from the DOI string, which this round forbids as a date source |
| Other milestones | |
| Latest update | · paper revision Basis: Preprint v2 posted Preprint v2 posted; the earlier event is the official blog and weights release (2026-02-18), which also published the developers' own multi-species task suite. Repository syncs through 2026-11-02 are maintenance Note: The latest update was recorded as the 2026-02-18 product update; corrected to the later paper revision source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Paper revision, weights release and repository maintenance within 12 months |
| Openness | Code: Open Weights: Research use only · custom licence (model hub label: other) Weights are on the model hub under a custom licence (checked 2026-11-16) with a JAX code release; the companion task suite is non-commercial |
| Benchmarks reported | 1 counted (self 1) |
| Wet-lab validation | 1 recorded campaign |
| Unverified fields | † wetlab_validation[0], open_status.code_license, background.funding |
| As of | 2026-11-16 |
What it does
Extends the two earlier generations from short-sequence classification to a 1 Mb context, with one model emitting functional tracks and gene-structure annotation and supporting conditional generation; post-training uses about 16,000 tracks and annotations across 24 species. It is far smaller than AlphaGenome but covers more species, and unlike Evo 2 it adds supervised track post-training.
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.
| Benchmark | Relation | Eval set / protocol | Source type | Note and evidence |
|---|---|---|---|---|
| 08 Basenji/Enformer test set | Developer report, official protocol | — | Self-reported | Self-report in a technical report, fine-tuned on re-folded splits · evidence |
Other evaluations (outside the 44)
- Developer-built multi-species task suite (Own task set) Self-reported: 106 tasks across 7 species with base-resolution outputs; released alongside the model, with competitors chosen by the developers and the win list paraphrased from the blog
- GUE / NT short-sequence suite (Public benchmark not among the 44) Self-reported: Not a pairing: the report does not use that suite and compares gene annotation on an earlier in-house benchmark instead; that benchmark entry's new version was released alongside this model's preprint
Developers report: one 1 Mb-context model covers track prediction, gene annotation and generation, beating several specialised tools and a larger sequence model on their own multi-species task suite.
Independent evidence: No independent evidence found. No third-party evaluation under a site benchmark protocol was found, and the in-house suite has no third-party re-run
Training data
Group in Figure 7: Public data + wet lab for validation only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Pretrained on a public pan-genome dataset (the report cites 9T bases trained) and post-trained on public functional-genomics tracks and annotations, including human and mouse track sets plus plant and insect data source |
| b. Private or licensed data | Not disclosed | |
| c. Own wet-lab data used for training | No | source |
| d. Wet lab used only for validation | Yes | Generated enhancers validated by reporter sequencing in one insect cell line, 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.
Denominator basis: 900 sequences = 600 activity-bin designs (60 per bin x 5 bins x 2 promoters) + a 150 vs 150 head-to-head; one cell type
The source is internally inconsistent (one section says 1,000, the methods say 900); 900 is used per the methods and the discrepancy noted. No hit rate is reported, only binning and promoter-specificity ratios
Executor: Developer-run and not peer reviewed
Background
- Organization: A decision-AI company founded in North Africa in 2014 and headquartered in London since 2015, with biology as a focus line; it became a subsidiary of an mRNA and immuno-oncology pharma company in 2023. At a 2025-10 event the developer said its nucleotide model family had over a million downloads and over five hundred citations. No citable official headcount, so none is given.
- Funding: A $7M Series A in 2019-05; a $100M Series B in 2022-01 (lead investor framings differ: one outlet says two funds co-led, another names an internet company); acquired by the parent in 2023-01 for about $389M upfront with up to about $243M in milestones, closing 2023-07-31.
- Products and partnerships: Builds on the family's two earlier nucleotide models (the previous generation was published in a journal in 2025-02); the developer also has a protein mass-spectrometry model, and the parent is an mRNA and immuno-oncology company. Three European and US research institutions collaborated. Comparators are Evo 2 and AlphaGenome.
- Related models: Evo 2, AlphaGenome, Borzoi
Reading notes
- The 106-task suite was built by the developers and released with the model, so it only shows leadership on tasks they defined.
- There is no GUE/NT suite result; on the track set only relative gains are given and the other big-lab model is not compared directly.
- The largest size is only 650M, so an edge over a smaller competitor configuration cannot be extrapolated to that competitor's largest one.
- The wet lab covers about 900 sequences in one cell type and reports binning and specificity ratios rather than hit rates, with the source inconsistent on the count.
Caveats
- The main evaluation is a developer-built suite with no third-party re-run.
- The technical report is not peer reviewed and the preprint has no journal version.
- The wet-lab count conflicts within the source (1,000 vs 900).
- Weights carry a custom licence and the suite is non-commercial.
Sources
- Technical report (preprint)
- Preprint version record (preprint)
- Official blog and weights release (company blog)
- Code repository (repository)
Corrections
3 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| first_release | 2025-12-22 (inferred from the DOI string) | 2025-12-25 (preprint v1 posted) | source | 2026-11-16 |
| latest_update | 2026-02-18, product update | 2026-10-29, preprint v2 (paper revision) | source | 2026-11-16 |
| wetlab_validation | 600 designs | 900 (600+150+150), noting that the source itself conflicts between 1,000 and 900 | 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.