Evo 2 (Arc Institute / NVIDIA)

Evo 2 is a biological foundation model — a language model over DNA rather than text — developed by the Arc Institute and NVIDIA with collaborators at Stanford University, UC Berkeley and UC San Francisco. It was trained on 9 trillion DNA base pairs from a curated genomic atlas spanning all domains of life, and has a 1-million-token context window at single-nucleotide resolution (Genome modelling and design across all domains of life with Evo 2). It is released fully open, including model parameters, training code, inference code and the training dataset. In August 2026 a Stanford team used Evo 2 to design complete bacteriophage genomes that were synthesized and shown to kill E. coli — reported as the first whole genomes designed by AI — which moved the model from a prediction and design tool into an active AI-biosecurity policy question.

Lineage

Evo 2 follows Evo, described in "Sequence modeling and design from molecular to genome scale with Evo" (Nguyen et al., Science 386, 2024; DOI 10.1126/science.ado9336), and a companion line of work on gene design published as "Semantic design of functional de novo genes from a genomic language model" (Merchant, King, Nguyen and Hie, Nature 649:749–758, 2026; DOI 10.1038/s41586-025-09749-7) (Genome modelling and design across all domains of life with Evo 2). Evo 2 was first released as a preprint in February 2025 and published in Nature the following year (Source: phys.org).

The paper carries more than sixty authors, with affiliations at the Arc Institute, NVIDIA, Stanford University, UC Berkeley, UC San Francisco, Columbia University, the University of Washington, Johns Hopkins University, Liquid AI, Goodfire and OpenAI. Brian L. Hie is the last author; Garyk Brixi is the first (Genome modelling and design across all domains of life with Evo 2).

Capabilities

The paper reports that Evo 2 predicts the functional impacts of genetic variation without task-specific fine-tuning, across a range from noncoding pathogenic mutations to clinically significant BRCA1 variants. Mechanistic interpretability analyses reported in the paper find that the model learns representations associated with biological features including exon–intron boundaries, transcription factor binding sites, protein structural elements and prophage genomic regions (Genome modelling and design across all domains of life with Evo 2).

On the generative side, the paper reports that Evo 2 produces mitochondrial, prokaryotic and eukaryotic sequences at genome scale "with greater naturalness and coherence than previous methods," and generates experimentally validated chromatin accessibility patterns when guided by predictive models and inference-time search (Genome modelling and design across all domains of life with Evo 2). The paper qualifies the genome-scale claim in its own text: the evaluation metrics "do not guarantee functional or replication-competent genomes," and the genome-scale generations "lack important elements, such as some essential genes" (Genome modelling and design across all domains of life with Evo 2).

The design result validated experimentally in the paper is epigenomic. Using predictive models and an inference-time beam search, the authors designed DNA with specified chromatin-accessibility profiles, synthesized it, integrated it site-specifically into mouse embryonic stem cell genomes, and measured accessibility by ATAC-seq. Three patterns encoding the Morse code messages "LO", "ARC" and "EVO2" reached experimental AUROCs of 0.92–0.95; 33 of 36 designs varying accessibility level within a sequence exceeded AUROC 0.8, while on the harder task of differential accessibility between two cell types 4 of 24 designs achieved greater than twofold differential accessibility (Genome modelling and design across all domains of life with Evo 2).

Reported prediction performance is uneven and the paper states the gaps: Evo 2 leads on ClinVar non-SNV coding and noncoding variants — insertions and deletions that models such as AlphaMissense and GPN-MSA cannot score — and on BRCA1 noncoding SNVs, but it underperforms state-of-the-art models on protein deep mutational scanning, remains modest on human gene essentiality (40B AUROC 0.66), and trails supervised sequence-to-function models such as ChromBPNet on distal regulatory variants (Genome modelling and design across all domains of life with Evo 2).

Genome design and laboratory validation

Stanford researchers led by Brian Hie, assistant professor of chemical engineering and Dieter Schwarz Foundation Stanford Data Science Faculty Fellow, with bioengineering graduate student Samuel H. King as first author, published "Generative design of bacteriophages with genome language models" in Science on August 6, 2026 (DOI 10.1126/science.aec2657). The paper's abstract describes the work as "the first generative design of complete bacteriophage genomes using genome language models." The team chose bacteriophage ΦX174, whose genome runs under 6,000 base pairs, as the starting point. Given a short seed of ΦX174 DNA, the model generated whole genomes end-to-end; Hie described the constraint as "we wanted the model to generate the entire genome end-to-end in a single left-to-right pass. We didn't add anything." The paper states that "nearly 300" of the generated genomes were chemically synthesized and tested, yielding 16 viable phages; the host organism is Escherichia coli C, and reporting of the same date gave the synthesis count as 302 (Generative design of bacteriophages with genome language models). Some designs showed higher fitness than the native ΦX174, and cryo-electron microscopy confirmed that one generated phage uses a DNA packaging protein from an evolutionarily distant phage in its capsid (Generative design of bacteriophages with genome language models; Source: news.stanford.edu). Press coverage described the result more broadly as the first complete genomes designed by AI (Source: theguardian.com); the paper's own claim is limited to bacteriophage genomes designed with genome language models (Generative design of bacteriophages with genome language models).

The stated motivation is resistance to antibiotic resistance. "If the bacteria gain resistance to a single phage, it's game over for the medication," Hie said. "But if you have multiple genetically distinct phages in a mixture, it would be harder for the bacteria to develop resistance to the entire cocktail." The paper reports a proof of concept in which a cocktail of the 16 phages overcomes resistance in E. coli that is immune to native ΦX174. King developed a computational framework to evaluate generated genomes against design criteria drawn from ΦX174 and related phages and narrow thousands of candidates to those worth synthesizing, DNA synthesis cost being the binding constraint (Source: news.stanford.edu).

Contributing authors include Stanford graduate students Claudia L. Driscoll, David B. Li, Daniel Guo, Aditi T. Merchant and Garyk Brixi, and Max E. Wilkinson of the Broad Institute of MIT and Harvard and Memorial Sloan Kettering Cancer Center. The work was funded by the Arc Institute, the National Science Foundation, the Knight-Hennessy Graduate Scholarship Fund, the Fannie and John Hertz Foundation, and the Stanford Institute for Human-Centered AI (Source: news.stanford.edu).

Availability

Evo 2's model parameters, training code, inference code and the OpenGenome2 training dataset were released openly, stated in the paper as intended "to accelerate the exploration and design of biological complexity" (Genome modelling and design across all domains of life with Evo 2). The code is published at github.com/arcinstitute/evo2 and the dataset on Hugging Face as arcinstitute/opengenome2; the model is also integrated into NVIDIA's BioNeMo framework under an Arc Institute–NVIDIA collaboration (Source: phys.org). This places Evo 2 among the open-weight releases whose dual-use profile is contested rather than governed by a developer-side deployment gate.

Safety and biosecurity debate

The paper reports one training-side control: genomic sequences from viruses that infect eukaryotic hosts were excluded from the training data for biosafety reasons (Genome modelling and design across all domains of life with Evo 2). Reporting on the August 2026 phage work describes the same exclusion as covering genetic code from viruses infecting humans, other animals or plants, to limit the risk of designing dangerous pathogens (Source: bbc.com). The authors state the exclusion's aim as ensuring "our openly shared model did not disseminate the capability to manipulate and design pathogenic human viruses," report two verifications that it took effect — high perplexity on eukaryotic viral sequences and no correlation between model likelihood and viral protein fitness — and state that red teaming meant to elicit pathogenic human viral proteins produced generations that were "effectively random in this domain." They also state its limit directly: "Task-specific post-training may circumvent this risk mitigation measure and should be approached with caution" (Genome modelling and design across all domains of life with Evo 2). The authors describe the release as aligned with the Responsible AI × Biodesign commitments and report data exclusion measures, safety and security evaluations, and population bias evaluations developed with multidisciplinary experts (Genome modelling and design across all domains of life with Evo 2). No system card or responsible-scaling-style framework accompanies the model.

Hie has argued that open availability is essential to expediting research and results. He acknowledges that modified versions of the tool could be used by bad actors, but argues that existing pathogens present a greater risk than potential AI designs because they are easier to access and produce, and that safety checks cannot be built into the process of natural evolution as they can into an AI tool; he further argues that tools such as Evo 2 give humans an advantage against naturally occurring pandemics and improved defense options against man-made biological threats (Source: news.stanford.edu).

Thomas V. Inglesby and Moritz S. Hanke of the Johns Hopkins Center for Health Security took the opposite position in a Science Perspective accompanying the phage paper, "AI-designed viral genomes," whose opening states that "the ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not" (AI-designed viral genomes). As quoted in reporting of the same date, they wrote that the findings raise "urgent biosafety and biosecurity questions" and that new viruses with the potential to cause disease "should not be pursued" (Source: theguardian.com). The Evo 2 paper itself cites the AI-biosecurity governance literature, including Bloomfield et al., "AI and biosecurity: The need for governance" (Science 385, 2024) (Genome modelling and design across all domains of life with Evo 2).

Open questions

• Hie has identified the model's own limits as unresolved: "how do we get greater genetic novelty and how do we get greater controllability of the outcomes?" (Source: news.stanford.edu).
• Whether the eukaryotic-virus training exclusion constrains a fine-tuned or modified open-weight derivative is not addressed in the sources reviewed here.
• Whether AI-designed genomes fall within existing US screening frameworks for synthetic nucleic-acid orders is not addressed in the sources reviewed here.

Related models

Evo 2 sits alongside other biology foundation models tracked in the same area: AlphaGenome (Google DeepMind) (Google DeepMind, regulatory variant-effect prediction, published in Nature January 2026) and AlphaFold (DeepMind) (protein structure). Where AlphaGenome is closed-weight and predictive, Evo 2 is open-weight and generative at genome scale, which is the axis on which its policy treatment differs.

Relationships

• developer: Arc Institute; Nvidia & TSMC — AI Compute Infrastructure
• related: AlphaGenome (Google DeepMind), AlphaFold (DeepMind)
• instance-of: Open-Weight Frontier Models
• related: AI Biosecurity, AI for Science, Dual-Use Frontier AI
• depends-on: Genome modelling and design across all domains of life with Evo 2
• supports: Generative design of bacteriophages with genome language models

Provenance note: Built from the Nature paper record and Arc Institute / Stanford releases, plus contemporaneous reporting on the August 6, 2026 Science paper. Both primary papers and the accompanying Perspective were ingested on August 7, 2026 and citations to them now resolve to Genome modelling and design across all domains of life with Evo 2, Generative design of bacteriophages with genome language models and AI-designed viral genomes. The Science paper and the Perspective are paywalled and were captured from publisher abstract records only.