Cheat sheet

gLM cheat sheet: Nucleotide Transformer v3

One reference card for Nucleotide Transformer v3: architecture, tokenization, size, context, training data, what it was shown to do and what it does not do. One card in a set that covers one genomic language model at a time.

GenomicsLLMMachine Learning

Genomic language model · Cheat sheet

Nucleotide Transformer v3

One backbone that reads DNA, predicts function and designs regulatory sequence

First public
bioRxiv, December 2025
Published
Preprint. Not yet peer reviewed.
Lab
InstaDeep, IMP Vienna, Cornell Tech, Cold Spring Harbor
Architecture
U-Net. A convolutional tower compresses the sequence, a Transformer bottleneck adds global context, and a matching tower expands it back.
Tokenization
Single nucleotide, single-base tokenization
Parameters
650M
Context
1,000,000 bases, with predictions restored at 1 base resolution.
Training data
9 trillion base pairs from OpenGenome2, over 128,000 species genomes, plus about 16,000 functional tracks and annotations across 24 species.
Used for
Reading DNA, predicting function and annotation, guided generation

Headline result

State of the art on the paper's own NTv3 Benchmark, 106 long-range tasks built to rule out overlap with training data, scored by correlation on functional tracks and MCC on annotation. Retrained to generate, the same backbone designed enhancers to a set activity level, and 1,000 were built and measured in cells.

What it does not do

The variant work shows its internals separate harmful from harmless expression variants, not that it outscores other variant tools. Still a preprint.

Commonly confused

The "60x" runs the other way from how it is repeated. The largest NTv3 model runs faster than competing models with up to 60x fewer parameters. Faster despite being bigger, not 60x smaller.

Download this card PNG, 1080 x 1350, sized for a post or a slide

Every model in the landscape figure is one dot with a tooltip. That is enough to compare them and not enough to use one. These cards are the other half: the same ten fields for every model, in the same order, so you can read one without having read any of the others.

The colour on the tokenization row is the same colour the model carries on the landscape chart, which makes each card a zoom-in on that figure rather than a separate thing to learn.

Two of the fields matter more than the rest. Headline result gives the metric and the condition it was measured under, because the number on its own does not tell you much. What it does not do is the field summaries usually leave out, and it is often the one that tells you whether the model fits your problem.

On NTv3 specifically, the “60x” figure runs the opposite way from how it usually gets repeated. The paper says the largest NTv3 model is faster at inference than competing models that have up to 60x fewer parameters. So NTv3 is faster in spite of being the bigger model, not 60x smaller than anything. It is a claim about the U-Net design paying for itself, and it is easy to invert by accident.

The design is what the card’s architecture row is describing. A convolutional tower compresses a megabase of sequence, a Transformer bottleneck reasons over the compressed view, and a matching tower expands back out, with skip connections carrying the local detail through. That is how it avoids running full attention at megabase length while still returning predictions at single-base resolution.

The third thing it does is design DNA, which no other model in this set does from the same weights. Fine-tuned by masked diffusion, without adding parameters, the same backbone generated enhancers with a specified activity level and promoter selectivity, and 1,000 of those sequences were tested by STARR-seq in cells.

One claim on the card is deliberately narrow. The variant work is an attribution analysis showing that predicted disruption is significantly larger for pathogenic variants on GTEx eQTLs. That is a finding about what the model is sensitive to, not a benchmarked variant effect scorer you can pick up and run.

This one is still a preprint.

Numbers are from the bioRxiv preprint. Sizes are the largest reported checkpoint, and release is the first public preprint, matching the conventions used on the landscape figure.