Multi-Modal Foundation · InstaDeepAI
NTv3 650M
A multi-species genomics foundation model predicting functional tracks and genomic elements across 1 Mb contexts.
What NTv3 650M is
NTv3 650M is a multi-species genomics foundation model created by InstaDeepAI. It addresses the computational challenge of capturing long-range regulatory context in DNA by processing sequences up to 1 megabase (1 Mb) in length at single-nucleotide resolution.
This supervised, post-trained variant builds on the pre-trained NTv3 foundation model, which was trained on 9 trillion base pairs. Using species-conditioning tokens, it predicts functional genomics tracks (in BigWig format) and annotated genomic elements (in BED format) across 24 animal and plant species alongside standard masked language modelling outputs.
How it works
The model employs a conditioned U-Net-like architecture featuring seven downsampling stages with adaptive layer normalisation, followed by a 12-layer transformer stack with 24 attention heads and multi-species prediction heads. DNA is processed using a character-level tokeniser with a vocabulary of 11 tokens covering the nucleotides A, T, C, and G, undetermined bases (N), and special tokens. A dedicated conditioning tokeniser encodes species and assembly identifiers.
Input sequences require padding to a multiple of 128 tokens. During inference, the network outputs masked language modelling logits, continuous BigWig functional genomics tracks, BED element predictions, and embeddings from both transformer and deconvolutional layers. Originally trained with bfloat16 mixed precision in JAX, the PyTorch implementation runs in float32 by default to preserve numerical stability, with optional bfloat16 execution on compatible hardware.
Key facts
- Architecture
- Conditioned U-Net with 7 downsample stages, 12-layer Transformer, and deconvolutional heads
- Context length
- Up to 1 Mb
- Input resolution
- Single nucleotide (character-level)
- Pre-training volume
- 9 trillion base pairs
- Species support
- 24 animal and plant assemblies
- Vocabulary size
- 11 tokens
What you can use it for
- Continuous functional track prediction
- Predicting multi-species continuous functional genomic signals in BigWig format across genomic intervals up to 1 Mb.
- Genomic element annotation
- Identifying discrete genomic elements in BED format across supported animal and plant assemblies.
- Masked language modelling
- Evaluating base-level likelihoods and sequence representations over masked DNA sequences.
- Multi-layer feature extraction
- Extracting intermediate representations from specific transformer layers, attention heads, or deconvolution stages for downstream analysis.
Run it here
We list this model but cannot run it yet: excluded (gated): Gated and non-commercial..
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Processes long-range sequence context up to 1 Mb at single-nucleotide resolution.
- Conditioned to make predictions across 24 animal and plant species.
- Produces both continuous BigWig functional tracks and discrete BED annotations simultaneously.
- Provides fine-grained control to extract embeddings and attention maps from specific intermediate transformer and deconvolution layers.
Watch out for
- Input sequence lengths must strictly be a multiple of 128.
- Conditioning is restricted to the 24 pre-configured species assemblies.
- Running in bfloat16 introduces minor numerical differences compared to the original JAX implementation.
- Requires execution of custom code and a transformers version of 4.55.0 or newer.
Inputs and outputs
You provide
Sequence with a mask
Use the model's mask token, commonly [MASK] or <mask>.
You get back
Labels with confidence scores, highest first.
Licence
Check the licence before commercial use · other
This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
Questions
Can I use NTv3 650M commercially?
Check the licence before commercial use. This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
What species does this model support?
The model incorporates conditioning tokens for 24 animal and plant species assemblies, including human and mouse.
What are the input formatting requirements?
Sequences must be tokenised at the character level across the canonical bases A, T, C, G, and N, with total sequence lengths padded to a multiple of 128.
What outputs does the model provide?
It provides masked language modelling logits, BigWig track logits, BED track logits, and embedding representations output by both the transformer and deconvolutional modules.
Can the model run in mixed precision?
Yes. It loads in full float32 precision by default, but compute datatypes across the stem, convolutions, and transformer layers can be configured to bfloat16 on supported GPUs.
Written from NTv3 650M's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.