Literature · marin-dna
MarinDNA m5.1 1B base model
A 1.12-billion-parameter causal language model for DNA sequence generation and variant effect prediction.
What MarinDNA m5.1 1B base model is
MarinDNA m5.1 is a 1.12-billion-parameter, nucleotide-level causal language model developed with Marin. It processes genomic sequences autoregressively at single-base resolution to assist researchers in evaluating genomic sequences and genetic variations.
The model was developed to assess standard Transformer architectures on genomic tasks, including variant effect prediction, using training data derived from diverse genomic elements across animal genomes.
How it works
The model employs a decoder-only Transformer architecture compatible with Qwen3 causal language models. It contains 1,120,772,224 parameters configured across 19 layers, with a hidden dimension of 1,920, an intermediate size of 7,680, and 15 attention heads matched with 15 key-value heads. The model accepts a context window of 256 tokens, which holds a beginning-of-sequence token followed by up to 255 DNA bases.
Training encompassed approximately 166.0 billion nucleotide tokens across a three-phase lineage. The model initially trained on a uniform three-region mixture of coding sequences (CDS), upstream regions, and downstream regions, followed by a continued run on the same mixture. The final stage introduced a uniform five-region mixture that added animal enhancer sequences and non-coding RNA (ncRNA) exons.
Key facts
- Architecture
- Qwen3-compatible decoder-only Transformer
- Parameters
- 1,120,772,224
- Layers
- 19
- Hidden size
- 1,920
- Intermediate size
- 7,680
- Attention heads
- 15 (15 key-value heads)
- Context length
- 256 tokens (1 BOS token and up to 255 DNA bases)
- Training exposure
- 166.0 billion nucleotide tokens
- Weight precision
- float32
What you can use it for
- Variant effect prediction
- Evaluating the functional impact of single-nucleotide alterations within short genomic sequences.
- Causal DNA generation
- Generating and completing synthetic or modified DNA sequences nucleotide by nucleotide.
- Genomic sequence modeling
- Analysing sequence patterns across coding sequences, upstream promoter elements, downstream intervals, enhancers, and non-coding RNA.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
MarinDNA m5.1 1B base model
Strengths and limits
Good at
- Incorporates training across five functional genomic regions: coding sequences, upstream, downstream, enhancers, and non-coding RNA.
- Uses single-nucleotide tokenisation to maintain individual base-pair resolution without subword fragmentation.
- Delivers competitive variant effect prediction performance within a standard 1.12-billion-parameter Transformer framework.
Watch out for
- Context window is restricted to a maximum of 255 DNA nucleotides due to the 256-token limit.
- Only recognises canonical DNA bases (A, C, G, T), mapping any ambiguous or modified IUPAC nucleotide characters to an unknown token.
- Lacks an end-of-sequence token, requiring users to define external stopping conditions for sequence generation.
Inputs and outputs
You provide
Prompt
The model continues from here.
You get back
The model's completion.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use MarinDNA m5.1 1B base model commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
How do I run MarinDNA m5.1 1B base model without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 16GB, typically in 12s, and a run costs about 70 credits. Visitors without an account get 1,000 credits a month.
What is the maximum input sequence length?
The model accepts up to 255 DNA bases per forward pass, as the first of its 256 context positions is reserved for an automatically prepended beginning-of-sequence token.
How are DNA sequences tokenised?
The model uses a seven-token, case-insensitive single-nucleotide vocabulary comprising pad, unknown, beginning-of-sequence, and the four canonical DNA bases (a, c, g, t).
What genomic datasets were used during training?
The training lineage used animal genomic datasets covering coding sequences, upstream regions, downstream regions, non-promoter cis-regulatory enhancer elements, and non-coding RNA exons.
Written from MarinDNA m5.1 1B base model'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.