Evo
Open genomic language models that read and write DNA, RNA and protein sequences
Evo is a family of open-source AI models built to handle genomic sequences, created by the Arc Institute with Stanford and UC Berkeley. Evo 1 arrived in February 2024 and Evo 2 followed in 2025, and the models have been used to design new CRISPR systems and complete phage genomes.
- Evo 1 opened with a 131,072 base pair window, against roughly 8,000 before
- Evo 2 holds 40 billion parameters and reads up to one megabase
- OpenGenome2 spans 8.8 trillion base pairs across all forms of life
- 16 AI-designed phage genomes proved viable in E. coli, some killing faster
- EvoCas9-1 cuts DNA like natural Cas9 at about 73% sequence similarity
What it is
Evo is a group of AI models built to work with genomic sequences, meaning the letter-by-letter code of DNA, RNA and proteins. Researchers at the Arc Institute, Stanford University and UC Berkeley created them, training the systems on DNA data. The first release, Evo 1 v.0.1.1, appeared on 27 February 2024. Evo 2, a 40-billion-parameter model, followed in 2025, and a smaller Evo 2 20B build came out on 28 February 2026 so it could run on one H100 GPU. The models are open source and are meant to analyse and generate biological sequences rather than text.
How it came about
Work on the line began at the Arc Institute together with Stanford and UC Berkeley. Evo 1 was published as version 0.1.1 in late February 2024. A larger successor, Evo 2, was introduced by the institute and its partners in 2025. The newest entry, Evo 2 20B, was released on 28 February 2026 as a trimmed-down variant that fits on a single H100 graphics card.
Inside the model
The system is described as a multi-modal genomic language model, meaning it handles DNA, RNA and protein sequences together rather than specialising in one task such as protein folding. Its backbone is the StripedHyena design, which combines state-space models with signal-processing operators to keep track of very long sequences. Evo 1 could attend to 131,072 base pairs at once, far beyond the roughly 8,000 typical of earlier genomic transformers. It learned from OpenGenome, about 300 billion nucleotides drawn from millions of prokaryote and phage genomes. Evo 2 was trained on OpenGenome2, 8.8 trillion base pairs spanning all life including eukaryotes, and stretched the window to one megabase at single-nucleotide resolution.
What it can do
Because it works across DNA, RNA and proteins at once, Evo can produce matching sequences of each. Researchers used it to invent new CRISPR systems, including EvoCas9-1, which cuts DNA about as well as natural Cas9 despite sharing only around 73% of its sequence. It also designed working transposons, the mobile elements that move DNA around a genome. With a million-base window, Evo 2 made zero-shot predictions about human non-coding variants, BRCA1 disease mutations and chromatin accessibility, cases it had not been trained on directly.
How to use it
All Evo models are released as open source. The software is written in Python, and the project page is hosted by the Arc Institute. The latest build, Evo 2 20B, was sized deliberately so a single H100 GPU is enough to run it.
Safety and testing
Early experiments kept viruses and pathogens that can infect people or complex organisms out of the training data. Lab work was limited to harmless bacterial hosts inside secure facilities. Safety studies published alongside the work proposed governance frameworks and screening protocols for companies that synthesise DNA. Even so, the 2026 Science paper on designed viral genomes was treated as a serious risk. Moritz Hanke of the Johns Hopkins Center for Health Security said producing functional viral genomes carries urgent biosafety and biosecurity implications, and pointed out that a new US National Institutes of Health policy on high-risk life sciences research covers known pathogens but not novel ones an AI might devise.
Reception and debate
The 2026 Science paper drew attention and unease in equal measure. It showed Evo generating hundreds of thousands of candidate bacteriophage genomes modelled on Phi X 174; when synthesised and tested on Escherichia coli, 16 of them, including Evo-Φ2147, proved viable, lysing and copying themselves inside the bacteria. Some killed bacteria faster than natural phages, which raised hopes for phage therapy against antimicrobial resistance, while commentators warned about the same capability being turned to harmful ends.
Why it matters
Evo broadened what genomic models could do, moving beyond single tasks such as gene expression or protein folding to whole-sequence design across DNA, RNA and proteins. Its outputs have fed into new CRISPR tools, synthetic transposons and engineered phages. The phage results also pushed discussions about screening synthetic DNA and governing AI that can write viral genomes.
Questions people ask about Evo
What is Evo for?
It is a family of models that read and write genomic sequences, covering DNA, RNA and proteins together. Rather than predicting one property such as protein shape, it analyses and generates biological code, and researchers have used it to design CRISPR systems, transposons and whole phage genomes.
Who built it and when did it appear?
The Arc Institute led the work with Stanford University and UC Berkeley. Evo 1 v.0.1.1 came out on 27 February 2024, Evo 2 followed in 2025, and a smaller Evo 2 20B version was released on 28 February 2026 to run on one H100 GPU.
How is it different from something like AlphaFold?
AlphaFold focuses on protein folding. Evo is a multi-modal genomic language model that handles DNA, RNA and proteins at once, using a StripedHyena design with state-space models so it can work over very long sequences, up to a megabase in Evo 2.
Can I use it myself?
Yes. The models are open source, written in Python, and published by the Arc Institute. The most recent release, Evo 2 20B, was cut down in size specifically so it fits on a single H100 GPU.
Why are people worried about it?
A 2026 paper in Science showed Evo producing complete genomes for new bacteriophage viruses, some of which worked better than natural ones. Experts said functional viral genomes raise urgent biosafety and biosecurity issues, and that current US policy on high-risk research covers known pathogens but not novel ones.
Good reading and sources
The facts on this page come from the Wikipedia article "Evo (AI)", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.