Compare · 24 models
DNA and genomics models you can run online
DNA and genomic foundation models process raw nucleotide sequences to predict regulatory elements, assess variant effects, and generate biological sequences across single loci or million-base-pair spans.
What dna and genomics models models do
Genomic foundation models address the challenge of reading unannotated DNA sequence to uncover regulatory logic, quantify functional phenotypes, and interpret the consequences of genetic variation. Rather than relying solely on hand-crafted genomic features, these architectures model raw sequence directly—either autoregressively as causal language models, bidirectionally via masked sequence modelling, or through specialised sequence-to-function convolutional and transformer backbones.
Across these systems, nucleotide sequences are typically encoded into individual bases or multi-nucleotide segments such as 6-mers and byte-pair encodings. While some models output continuous predictive tracks for molecular features such as chromatin accessibility, transcription factor binding, and gene expression, others generate nucleotide likelihoods or dense sequence embeddings that support downstream zero-shot variant scoring, generation, or fine-tuning across diverse phylogenetic clades.
Compared
- Evo 2Arc Institute / StanfordA genome-scale foundation model for nucleotide sequence generation, variant effect scoring, and genome modelling.Check termsReady
- Nucleotide Transformer v2InstaDeep / Wellcome SangerA DNA foundation model that generates nucleotide sequence representations for regulatory element and variant analysis.Non-commercialReady
- Carbon-3BHuggingFaceBioA 3-billion-parameter autoregressive DNA foundation model for sequence generation and variant-effect prediction.Check termsOn demand
- Carbon-500MHuggingFaceBioA 500M-parameter autoregressive DNA language model designed primarily as a draft model for speculative decoding.Check termsOn demand
- Carbon-8BHuggingFaceBioAn 8B-parameter autoregressive DNA model for genomic prediction, generation, and long-context sequence retrieval.Check termsOn demand
- ChatNTInstaDeepAIConversational model answering natural-language queries for DNA, RNA, and protein sequence tasks.Check termsOn demand
- EnformerDeepMindPredicts gene expression and regulatory tracks across long genomic windows to assess variant effects.Commercial OKOn demand
- GENA-LM BERT base (T2T)AIRI-InstituteA BERT-base genomic foundation model trained on T2T human assemblies for masked DNA sequence understanding.Check termsOn demand
- GENA-LM BERT large (T2T)AIRI-InstituteGenomic foundation model based on BERT large, trained on the T2T human genome assembly for sequence analysis.Check termsOn demand
- Genos 1.2BZhejiangLabA 1.2-billion-parameter genomic foundation model for high-efficiency sequence analysis and prediction tasks.Check termsOn demand
- gLM2 650MtattabioA 650M-parameter genomic foundation model encoding contigs with both amino acid and nucleotide tokens.Check termsOn demand
- MarinDNA m5.1 1B base modelmarin-dnaA 1.12-billion-parameter causal language model for DNA sequence generation and variant effect prediction.Check termsOn demand
- ModernGENA baseAIRI-InstituteA 377M-parameter ModernBERT DNA foundation model for genomic sequence modelling and downstream analysis tasks.Check termsOn demand
- ModernGENA largeAIRI-InstituteA 377M-parameter ModernBERT foundation model for vertebrate DNA sequence modelling and masked sequence prediction.Check termsOn demand
- Nucleotide Transformer 2.5BInstaDeepAIA 2.5B-parameter DNA language model trained across 850 species genomes for molecular phenotype prediction.Check termsOn demand
- OneGenome-RiceZhejiangLabA genomic foundation model analysing long DNA sequences for rice functional genomics and precision breeding.Check termsOn demand
- AlphaGenomegooglePredicts DNA regulatory features, such as gene expression and TF binding, at single-nucleotide resolution.See its pageNot yet
- Botanic0-Lliving-modelsPlant genomic foundation model for regulatory annotation, expression analysis, and variant effect prediction.See its pageNot yet
- Botanic0-Mliving-modelsA mid-sized plant genomic foundation model for regulatory annotation, expression inference, and variant effect analysis.See its pageNot yet
- Botanic0-Sliving-modelsA compact 4-layer plant genomic foundation model pre-trained on nuclear DNA across 43 diverse plant genomes.See its pageNot yet
- Botanic1-Lliving-modelsPlant genomic foundation model for sequence embedding, zero-shot variant effect scoring, and downstream fine-tuning.See its pageNot yet
- Botanic1-Mliving-modelsA 688M-parameter bidirectional state-space plant genomic foundation model for sequence modelling.See its pageNot yet
- Botanic1-M SAE (k=64, codebook 8192)living-modelsSparse autoencoders decomposing Botanic1-M activations into 8,192 features for plant genomic interpretability.See its pageNot yet
- Botanic1-Sliving-modelsA 318M-parameter bidirectional Mamba-2 foundation model for analysing plant genomes at single-nucleotide resolution.See its pageNot yet
How to choose
Selection depends primarily on sequence context requirements, architectural objective, and the biological taxa under study. For broad multi-kingdom or eukaryotic sequence generation and likelihood-based variant scoring across large windows, Evo 2 and the Carbon series provide deep context windows. Evo 2 scales context up to 1 million tokens in its 7B and 40B checkpoints across the tree of life, but demands 80 GB of GPU memory. In contrast, Carbon-8B and Carbon-3B target eukaryotic genomes using 6-mer tokenisation and Factorised Nucleotide Supervision for single-base scoring. Carbon-3B reaches generation speeds beyond 100,000 base pairs per second on an H100 with vLLM and handles contexts up to approximately 197 kilobase pairs, natively pairing with Carbon-500M for speculative decoding. When computational resources are constrained to consumer hardware, Carbon-500M requires only 16 GB of GPU memory for its 8,192-token context window.
For regulatory element classification and molecular phenotype prediction from sequence, bidirectional encoder models are standard. Nucleotide Transformer v2 (500M parameters) and Nucleotide Transformer 2.5B learn sequence representations from 850 diverse genomes across 2,048-token and 1,000-token windows respectively, but exclude plant and viral genomes. If the focus is human-specific genomics, GENA-LM BERT base and large (T2T) are pre-trained on complete Telomere-to-Telomere human assemblies augmented with gnomAD single-nucleotide polymorphisms, compressing roughly 4,500 nucleotides into 512 byte-pair tokens. For modernized encoder efficiency, ModernGENA base and ModernGENA large adapt ModernBERT with hybrid attention and FlashAttention compatibility, though their pre-training concentrates on transcription start site regions rather than unbiased whole-genome tiling.
When directly predicting experimental functional tracks rather than masked tokens, Enformer models regulatory interactions up to 100 kilobases away via a 200-kilobase receptive field, producing simultaneous binned predictions for CAGE, DNase sensitivity, and histone marks. AlphaGenome similarly targets multi-modal regulatory prediction, including gene expression, chromatin accessibility, and transcription factor binding, resolving features at single-nucleotide precision.
For specialised biological domains, standard general-purpose models often fall short. Plant research is served by OneGenome-Rice—a 1.25-billion-parameter Mixture-of-Experts architecture processing 1 million base pairs across 422 Oryza accessions—and the Botanic series. Botanic0-S, Botanic0-M, and Botanic0-L span 4 to 40 transformer layers trained on 43 plant genomes, while the Botanic1 family (Botanic1-S, Botanic1-M, and Botanic1-L) uses bidirectional Mamba-2 state-space blocks without attention, scaling up to 2.1 billion parameters; Botanic1-M SAE provides post-hoc interpretability via 8,192-feature autoencoders. For mixed-modality tasks, gLM2 650M natively encodes contigs combining uppercase amino acid coding sequences and lowercase nucleotide inter-genic spans, while ChatNT combines a Nucleotide Transformer v2 encoder with natural-language dialogue across DNA, RNA, and protein instructions.
Questions
Can I run these models without a GPU or installing anything?
16 of the 24 models here run on our servers from their page: open one, fill in the inputs and press Run. Those marked "Not yet" are listed with their licence and documentation but have no loader here yet.
Which models support sequence contexts of up to one million base pairs?
Evo 2 (in its 7B and 40B parameter versions), Genos 1.2B, and OneGenome-Rice natively support sequence context windows reaching one million base pairs or tokens.
Which tools are specifically designed for plant genomic sequences?
OneGenome-Rice is specialised for rice (Oryza sativa) functional genomics and breeding. The Botanic0 series (Botanic0-S, Botanic0-M, Botanic0-L) and Botanic1 series (Botanic1-S, Botanic1-M, Botanic1-L) are dedicated plant models pre-trained across diverse embryophyte genomes.
How do tokenisation methods differ across DNA foundation models?
Models take varying approaches: MarinDNA m5.1 and Botanic1-S use single-nucleotide tokenisation for direct base-pair resolution; Nucleotide Transformer v2, Nucleotide Transformer 2.5B, and Botanic0-S segment DNA into 6-mers; the GENA-LM BERT models use byte-pair encoding; and the Carbon family uses a hybrid vocabulary combining non-overlapping 6-mers with text BPE alongside factorised nucleotide supervision.
Can I query genomic sequences using natural language instructions?
Yes, ChatNT is an 8-billion-parameter multimodal instruction-tuned model that accepts natural-language queries across 27 sequence tasks spanning DNA, RNA, and proteins.
Written from each model's own documentation, as summarised on its page here. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.