[!TIP] The MultiMolecule team has confirmed that the provided model and checkpoints are producing the same intermediate representations as the original implementation.

The team releasing HyenaDNA did not write this model card for this model so this model card has been written by the MultiMolecule team.

Model Details

HyenaDNA is a decoder-only model pre-trained on the human reference genome with single nucleotide tokenization in a self-supervised fashion. This means that the model was trained on the raw nucleotides of DNA sequences only, with an automatic process to generate inputs and labels from those sequences. Please refer to the Training Details section for more information on the training process.

HyenaDNA replaces the attention mechanism with the Hyena operator — a subquadratic sequence mixer based on implicit long convolutions. This enables O(L log L) complexity for sequence modeling, allowing the model to handle context lengths up to 1 million base pairs at single nucleotide resolution.

The Hyena operator uses:

Implicit filters: MLP-parameterized convolution kernels with learned positional embeddings
Element-wise gating: Multiplicative interactions between projected input channels
FFT convolution: Efficient computation via the Fast Fourier Transform

Variants

HyenaDNA was pretrained across 5 architectures (Table A.1 in the paper), of which 4 have released checkpoints. We provide one converted checkpoint per architecture, using the longest available context length:

multimolecule/hyenadna-large: The HyenaDNA large model with 8 layers, 256 hidden size, and 1,000,000 context length.
multimolecule/hyenadna-medium: The HyenaDNA medium model with 4 layers, 256 hidden size, and 32,768 context length.
multimolecule/hyenadna-small: The HyenaDNA small model with 2 layers, 256 hidden size, and 1,024 context length.
multimolecule/hyenadna-tiny: The HyenaDNA tiny model with 2 layers, 128 hidden size, and 16,384 context length.

[!NOTE] The original repository releases 7 checkpoints with inconsistent naming (e.g., "large" and "medium" share the same architecture, and multiple "tiny" variants are identical except for context length). We retain only the best (longest context) checkpoint per architecture and use a consistent size-based naming scheme. The mapping from original to MultiMolecule names is:

Original Checkpoint MultiMolecule Name Architecture
hyenadna-large-1m-seqlen-hf hyenadna-large 8 layers, 256 dim
hyenadna-small-32k-seqlen-hf hyenadna-medium 4 layers, 256 dim
hyenadna-tiny-1k-seqlen-d256-hf hyenadna-small 2 layers, 256 dim
hyenadna-tiny-16k-seqlen-d128-hf hyenadna-tiny 2 layers, 128 dim

The (4, 128) architecture from the paper has no released checkpoint. The remaining 3 original checkpoints (hyenadna-medium-450k-seqlen-hf, hyenadna-medium-160k-seqlen-hf, hyenadna-tiny-1k-seqlen-hf) share the same architecture as one of the above but with shorter context lengths, and are not provided.

Model Specification

Variants	Num Layers	Hidden Size	Intermediate Size	Num Parameters (M)	FLOPs (G)	MACs (G)	Max Num Tokens
HyenaDNA-large	8	256	1024	6.62	6.69	3.35	1,000,002
HyenaDNA-medium	4	3.34	3.35	1.67	32,770
HyenaDNA-small	2	1.71	1.67	0.84	1,026
HyenaDNA-tiny	128	512	0.45	0.44	0.22	16,386

Usage

The model file depends on the multimolecule library. You can install it using pip:

pip install multimolecule

Direct Use

Text Generation

You can use this model directly with a pipeline for text generation:

python

import multimolecule  # you must import multimolecule to register models
from transformers import pipeline

generator = pipeline("text-generation", model="multimolecule/hyenadna-large")
output = generator("ATCGATCGATCG", max_new_tokens=50)

Downstream Use

Extract Features

Here is how to use this model to get the features of a given sequence in PyTorch:

python

from multimolecule import DnaTokenizer, HyenaDnaModel


tokenizer = DnaTokenizer.from_pretrained("multimolecule/hyenadna-large")
model = HyenaDnaModel.from_pretrained("multimolecule/hyenadna-large")

text = "ATCGATCGATCGATCG"
input = tokenizer(text, return_tensors="pt")

output = model(**input)

Sequence Classification / Regression

[!NOTE] This model is not fine-tuned for any specific task. You will need to fine-tune the model on a downstream task to use it for sequence classification or regression.

Here is how to use this model as backbone to fine-tune for a sequence-level task in PyTorch:

python

import torch
from multimolecule import DnaTokenizer, HyenaDnaForSequencePrediction


tokenizer = DnaTokenizer.from_pretrained("multimolecule/hyenadna-large")
model = HyenaDnaForSequencePrediction.from_pretrained("multimolecule/hyenadna-large")

text = "ATCGATCGATCGATCG"
input = tokenizer(text, return_tensors="pt")
label = torch.tensor([1])

output = model(**input, labels=label)

Token Classification / Regression

[!NOTE] This model is not fine-tuned for any specific task. You will need to fine-tune the model on a downstream task to use it for token classification or regression.

Here is how to use this model as backbone to fine-tune for a nucleotide-level task in PyTorch:

python

import torch
from multimolecule import DnaTokenizer, HyenaDnaForTokenPrediction


tokenizer = DnaTokenizer.from_pretrained("multimolecule/hyenadna-large")
model = HyenaDnaForTokenPrediction.from_pretrained("multimolecule/hyenadna-large")

text = "ATCGATCGATCGATCG"
input = tokenizer(text, return_tensors="pt")
label = torch.randint(2, (len(text), ))

output = model(**input, labels=label)

Training Details

HyenaDNA used Causal Language Modeling (CLM) as the pre-training objective: given a DNA sequence, the model is trained to predict the next nucleotide token autoregressively.

Training Data

The HyenaDNA model was pre-trained on the human reference genome (GRCh38). The training data consists of single nucleotide-level DNA sequences from all human chromosomes. Sequences are tokenized at the individual character level (A, C, G, T, N) without k-mer encoding.

The dataset is split into training and test sets by chromosome, with held-out chromosomes used for evaluation.

Training Procedure

Preprocessing

HyenaDNA used causal language modeling (CLM) as the pre-training objective: given a DNA sequence of length L, the model predicts the next nucleotide at each position, i.e., predicting token x_{t+1} given x_1, ..., x_t.

Single nucleotide tokenization is used with a vocabulary of 12 tokens: A, C, G, T, N, and special tokens (CLS, SEP, BOS, MASK, PAD, RESERVED, UNK).

Sequence Length Warm-up

A key training strategy in HyenaDNA is progressive sequence length warm-up. Training begins with short sequences and gradually increases the context length:

Training starts with sequences of length L=64.
The sequence length is doubled at each warm-up stage (64 → 128 → 256 → ... → target length).
This strategy enables stable training at very long context lengths that would be difficult to train from scratch.

Pre-training

The model was trained on up to 8 NVIDIA A100 (80GB) GPUs.

Batch size: 64 -- 256
Steps: 10,000 -- 20,000
Optimizer: AdamW
Learning rate: 1.5e-4 -- 6e-4
Learning rate scheduler: Cosine
Weight decay: 0.1

Citation

bibtex

@inproceedings{nguyen2023hyenadna,
  title={Hyena{DNA}: Long-Range Genomic Sequence Modeling at Single Nucleotide Resolution},
  author={Eric Nguyen and Michael Poli and Marjan Faizi and Armin W Thomas and Michael Wornow and Callum Birch-Sykes and Stefano Massaroli and Aman Patel and Clayton M. Rabideau and Yoshua Bengio and Stefano Ermon and Christopher Re and Stephen Baccus},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023},
  url={https://openreview.net/forum?id=ubzNoJjOKj}
}

[!NOTE] The artifacts distributed in this repository are part of the MultiMolecule project. If MultiMolecule supports your research, please cite the MultiMolecule project as follows:

bibtex

@software{chen_2024_12638419,
  author    = {Chen, Zhiyuan and Zhu, Sophia Y.},
  title     = {MultiMolecule},
  doi       = {10.5281/zenodo.12638419},
  publisher = {Zenodo},
  url       = {https://doi.org/10.5281/zenodo.12638419},
  year      = 2024,
  month     = may,
  day       = 4
}

Contact

Please use GitHub issues of MultiMolecule for any questions or comments on the model card.

Please contact the authors of the HyenaDNA paper for questions or comments on the paper/model.

License

This model implementation is licensed under the GNU Affero General Public License.

For additional terms and clarifications, please refer to our License FAQ.

SPDX-License-Identifier: AGPL-3.0-or-later

Original Checkpoint	MultiMolecule Name	Architecture
`hyenadna-large-1m-seqlen-hf`	`hyenadna-large`	8 layers, 256 dim
`hyenadna-small-32k-seqlen-hf`	`hyenadna-medium`	4 layers, 256 dim
`hyenadna-tiny-1k-seqlen-d256-hf`	`hyenadna-small`	2 layers, 256 dim
`hyenadna-tiny-16k-seqlen-d128-hf`	`hyenadna-tiny`	2 layers, 128 dim

multimolecule/hyenadna-large

HyenaDNA

Disclaimer

Model Details

Variants

Model Specification

Links

Usage

Direct Use

Text Generation

Downstream Use

Extract Features

Sequence Classification / Regression

Token Classification / Regression

Training Details

Training Data

Training Procedure

Preprocessing

Sequence Length Warm-up

Pre-training

Citation

Contact

License

multimolecule/
hyenadna-large