Evo is a new AI model built for biological sequence modeling, and its ambition is broader than a single protein task or narrow genomic prediction. Presented by a team from TogtherAI and the Arc Institute, it is designed to interpret DNA, RNA, and proteins while supporting generative design from the molecular level up to the genomic level.
The model was developed by Eric Nguyen, Michael Poli, Matthew Durrant, Patrick Hsu and Brian Hie. It uses a modified version of the StripedHyena architecture, which the team applies to one of the hardest problems in biological AI: handling very long sequences while still paying attention to tiny changes at the nucleotide level.
Why Evo matters for biological sequence modeling
Biological data is not short. DNA sequences can stretch to extreme lengths, while meaningful changes can occur at the level of a single nucleotide. That creates a difficult combination for AI systems: they need long-range context, but they also need high-resolution sensitivity.
Evo is built to work at the nucleotide level, recognizing and interpreting the smallest building blocks of DNA and RNA. The architecture can model long contexts and process more than 650,000 tokens. Evo can process sequences up to 131 kilobases (131,000 bases) in length.
That scale is central to the model’s purpose. A system that only sees short fragments may miss broader biological patterns. A system that compresses too aggressively may lose the fine-grained information needed to understand the effects of small mutations. Evo is presented as an attempt to bridge those demands inside one foundation model for biology.
"Evo tries to show a path forward toward unified and foundation modeling on biology," says Michael Poli, co-author of Evo and StripedHyena.
How the model was trained
Like language models, Evo uses a next-token prediction objective. In this case, the prediction happens at the nucleotide level rather than over ordinary text. The comparison matters because the team is applying a familiar foundation-model training idea to biological sequences, where the units, scale, and consequences are very different.
Poli describes the obstacle directly: "The problem up until now, why this hasn't been done, is that sequences are extremely long if you want to capture meaningful properties about DNA and also learning at high resolution is quite challenging for transformers," says Poli.
The source article explains that tokenizers can create problems in language models because they often do not work at the character level. Evo’s nucleotide-level approach is intended to keep the model close to the biological sequence itself.
Evo was trained on a large database of 2.7 million prokaryotic genomes, described as a fraction of the publicly available genomic data. Training happened in two stages:
- First, the model was trained with a context length of 8,000 base pairs.
- Second, the context length was increased to 131,000 base pairs.
The corresponding training dataset, OpenGenome, will be made publicly available shortly. According to the source, this longer context lets Evo recognize patterns and make predictions across longer DNA sequences than previous methods.
What Evo can predict
Early experiments show several possible applications for Evo. One is predicting an organism's vital genes based on small DNA mutations. The team says this capability could replace traditional laboratory experiments, which can often take months.
In tests, Evo was able to compete with leading protein-specific language models when predicting the effects of mutations on the function of E. coli proteins. That is important because Evo is not described as a protein-only system. Its broader claim is that one model can work across DNA, RNA, and proteins.
The model can also predict the functional properties of non-coding RNAs (ncRNAs). In addition, it can infer gene expression from regulatory DNA. Together, these examples show why the model is being positioned as a unified system rather than a tool for one biological modality.
The team also compared different architectures during training, including Transfomer models and Mamba. Poli argues that deep signal processing architectures performed especially well: "Well, the amazing thing is that these deep signal processing architectures seem to scale better," Poli says. "It's not just that they can process these longer sequences and then do about as well as transformers. It's as if they scale better per flop. They're just better architectures, I believe, than transformers."
Generative design at molecular and genomic scale
Evo is not only a prediction model. The team also presents it as a generative design system for biology. It can generate complex molecular systems such as CRISPR-Cas complexes and transposable elements.
The model can also generate DNA sequences longer than 650 kilobases, described as an order of magnitude larger than previous methods. While previous generative models typically focus on a single modality, Evo is capable of designing large functional complexes of proteins and ncRNAs.
That combination is the central claim around Evo: it can operate across biological languages and across biological scale. It can deal with nucleotide-level resolution, but it can also move toward larger systems that involve multiple kinds of biological components.
The safety questions around genomic-scale AI
The Evo team sees the model as a potential milestone in biological sequence modeling. The source names possible applications in chemistry, materials science, drug discovery, agriculture, and sustainability. At the same time, the team says practical use of generated sequences will require further validation.
The safety concern is direct. Evo is described as the first system of its kind that can predict and generate DNA sequences at the level of the entire genome, with single-nucleotide resolution. That capability creates both scientific opportunity and responsibility.
"Future capabilities that emerge from large-scale DNA models like Evo also require additional work to ensure that these capabilities are deployed safely and for the benefit of humanity," the blog post states.
The source identifies concerns about potential misuse, social and health injustice, and environmental degradation. The team suggests comprehensive guidelines for ethical practices, greater transparency, and international collaborations and partnerships for responsible use and development.
Investment in education and capacity building is also mentioned, along with collaboration with organizations such as the Global Alliance for Genomics and Health (GA4GH). The goal described in the source is a future where advances in genetic engineering remain aligned with ethical principles and societal values.
The team provides code and model via GitHub. For now, Evo’s importance lies in the direction it points: biological foundation models that can read, predict, and generate across DNA, RNA, and proteins, while forcing a parallel conversation about validation, governance, and responsible deployment.