Genomics & Single-Cell · Arc Institute / Stanford
Evo 2
A genome-scale foundation model for nucleotide sequence generation, variant effect scoring, and genome modelling.
What Evo 2 is
Evo 2 is a genomic foundation model developed by the Arc Institute and Stanford. It is designed to model genetic sequences across the tree of life, operating from individual nucleotides up to whole-genome loci.
The model is capable of both generating new sequences and scoring existing ones. Scientists can use its likelihood scores to evaluate variant effects, or generate sequence continuations to design biological features such as regulatory elements.
How it works
Evo 2 is an autoregressive DNA language model trained on trillions of tokens drawn from multi-kingdom DNA across the tree of life. It receives nucleotide sequences as input and produces sampled sequence continuations alongside logit vectors that reflect model sequence likelihoods.
The model family includes architectures structured across 25, 32, and 50 layers, scaling to 1 billion, 7 billion, and 40 billion parameters. Base checkpoints are trained on sequence context windows of 8,192 tokens, whereas extended 7B and 40B checkpoints scale up to 1 million sequence tokens.
Key facts
- Version
- 2.0
- Input modality
- Nucleotide sequence
- Parameter count
- 1B, 7B, and 40B
- Layer count
- 25 (1B), 32 (7B), 50 (40B)
- GPU memory
- 80 GB
What you can use it for
- Variant effect prediction
- Score potential mutational consequences by evaluating logit vectors and model likelihoods across DNA sequences.
- Regulatory element design
- Sample autoregressive continuations from nucleotide inputs to design new synthetic regulatory sequence elements.
- Genome locus modelling
- Analyse extended genomic contexts and whole loci using context windows reaching up to 1 million tokens.
- Nucleotide sequence generation
- Produce continuation sequences from arbitrary DNA prompts to explore sequence space across diverse organisms.
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 Managed GPU endpoint
Runs on managed compute. Availability and startup time depend on current capacity.Run this model
Evo 2
Strengths and limits
Good at
- Supports long sequence contexts of up to 1 million tokens for genomic-scale analysis in the 7B and 40B models.
- Trained across the tree of life on multi-kingdom DNA data totalling trillions of tokens.
- Provides both autoregressive sequence generation and output likelihoods usable as variant effect scores.
Watch out for
- Demands substantial computing hardware, requiring 80 GB of GPU memory.
- Base checkpoints possess a much smaller context window restricted to 8,192 tokens.
Inputs and outputs
You provide
DNA sequence
Nucleotide sequence (A/C/G/T).
Tokens to generateoptional
You get back
Continuation sampled from the model.
Model likelihood, usable as a variant effect score.
Licence
Check the licence before commercial use · Arc Institute License
The Arc Institute License licence carries terms worth reading before you depend on it.
Questions
Can I use Evo 2 commercially?
Check the licence before commercial use. The Arc Institute License licence carries terms worth reading before you depend on it.
How do I run Evo 2 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 2× A100, typically in 30s, and a run costs about 840 credits. Visitors without an account get 1,000 credits a month.
What sequence lengths can Evo 2 process?
The full 7B and 40B checkpoints support sequence lengths of up to 1 million tokens, whereas the 1B, 7B, and 40B base checkpoints support context windows of 8,192 tokens.
What outputs does the model produce from an input DNA sequence?
The model outputs generated continuation sequences based on requested token counts, as well as logit vectors indicating sequence likelihoods.
What parameter configurations are available?
Available models include a 25-layer 1B parameter checkpoint, a 32-layer 7B parameter checkpoint, and a 50-layer 40B parameter checkpoint.
What hardware is required to run the model?
Running the model requires 80 GB of GPU memory.
How to cite it
Written from Evo 2's own documentation. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.