NTv3 Foundation Models for Genomics

NTv3 Foundation Models for Genomics

Introduces InstaDeepAI's NTv3 genomic foundation models, which read up to 1 Mb of DNA, and points to runnable PyTorch notebooks, in a browser page you can also download.

Made by InstaDeepAI on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: NTv3 Foundation Models for Genomics is a companion hub from InstaDeepAI for a family of genomic models. The models take up to 1 Mb of DNA in one pass and give a prediction for every base. A single design handles three tasks: it predicts functional tracks, annotates genes and designs regulatory sequences. The Space is a static page that points to runnable PyTorch notebooks.

Who it suits: Genomics researchers and machine learning people who want to read about the models and run the tutorials.

Industry | Who uses it | How they use it | Example
Genomics research | Computational biologists | Predict functional tracks across a long stretch of DNA | Running a tracks model on a genome region
Plant biotechnology | Plant scientists | Design plant core promoters at a requested strength with the example generative model | Following notebook 06
Bioinformatics | Annotation teams | Predict genes from DNA sequence alone, with one model per clade | Picking the plants or the insects model

Skip it if you want a point-and-click web tool, since this Space is a hub of pages and notebooks and the models live in their own repositories.

Where to use it: The hub works in these places.

• A web browser: the static Space opens there with nothing to install
• Windows, macOS or Linux: download the Space's files with Python and open index.html
• PyTorch notebooks: the notebooks_tutorials folder holds the runnable ones

Use case | Who | What they get back
Read about the models | Researchers | The Space's pages on what NTv3 does and how its models differ
Run the tutorials | Developers | Runnable PyTorch notebooks from the notebooks_tutorials folder
Design a plant promoter | Plant scientists | Core promoter designs at a requested strength from notebook 06

How to set it up: Browsing the hub needs only a web browser. A local copy and the notebooks need a few more things.

• A web browser
• Python on Windows, macOS or Linux, to download the files
• A PyTorch setup, to run the notebooks

1. Open the Space's page on Hugging Face to browse it online.
2. For a local copy, install the Hugging Face hub library with pip.
3. Download the Space's files into a folder called ntv3.
4. Move into that folder.
5. Open index.html in a web browser, and find the notebooks in notebooks_tutorials.

pip install -U huggingface_hub
hf download InstaDeepAI/ntv3 --repo-type space --local-dir ntv3
cd ntv3

How to use it: Open the page and read how the models differ. Then open a notebook and run it with PyTorch. Notebook 05 teaches the recipe for building a design model, and notebook 06 runs the plant promoter model.

cd ntv3

Tool | What it does
NTv3-pretrained | Gives embeddings and masked-token guesses, and can be fine-tuned on your own data
NTv3-tracks | Predicts functional tracks for inputs between 16 kb and 1 Mb, and is the starting point for later fine-tuning
NTv3-tracks-1mb | A variant built for 1 Mb windows and used for the paper's results, so shorter windows are not supported
NTv3-annotate | Predicts genes from DNA sequence alone, with one model per clade
NTv3-generative-plant-promoter | A sample generative model that creates plant core promoters for the strength you choose

Example: You open notebook 06 from the notebooks_tutorials folder. It runs the plant promoter model and designs a core promoter at the strength you ask for.

For newcomers and experts: If you are new, NTv3 is a family of models that read DNA and predict what its parts do. Open the Space's page first, read the overview, then open one notebook.

If you know the field, the network follows a U-Net layout. Convolutions first read the sequence, Transformer layers pass information across long distances, and deconvolutions bring the output back down to single bases. Training has four stages. A self-supervised stage predicts masked tokens on OpenGenome2 genomes, a supervised stage learns tracks, a third fine-tunes annotation for each clade, and the last applies masked diffusion to MPRA activity data. Each released checkpoint is about 700M parameters.

Up to 1 Mb of DNA goes in
A convolutional tower reads the sequence
A Transformer tower carries long-range context
A deconvolutional tower returns per-base output

Good to know: Only some of the paper's models are released, and one of them is an example rather than a full tool.

• Design models: just the plant promoter model has been published, and the enhancer models for plants and flies are held back
• Species: the plant promoter example is not species-aware, so read its model card before quoting it
• Named on Hugging Face: six model repositories and four datasets
• Size: 37 files in the repository
• License: apache-2
• Cost to run: none named
• Data it reaches: the models and datasets it names on Hugging Face
• Not for: no forbidden uses are named, but the page has an Access section to read before you download weights

Title: Functional tracks by type
Gene expression: 9603 tracks
Histone marks: 3115 tracks
TF binding: 2193 tracks
Accessibility: 1351 tracks

On Hugging Face, 13 November 2026

• Likes: 38
• Built with: static
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/InstaDeepAI/ntv3
• Files: https://huggingface.co/spaces/InstaDeepAI/ntv3/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.