Skip to content
ContentLora

    Tip: press / anywhere to search.

    technology

    Evo 2

    Also known as Evo2, Evo 2 40B, Evo 2 7B

    Evo 2 is a genome language model from the Arc Institute and NVIDIA, released with a preprint on 19 February 2025 and published in Nature on 4 March 2026.[1][2] It was trained on about 9 trillion DNA base pairs from all domains of life and reads up to 1 million tokens at single-nucleotide resolution.[2] Its developers released the training data, code and model weights openly.[3]

    Editor reviewedUpdated Wiki13 claims · 100% primaryAI for science, Synthetic biology, Life sciences
    Key facts

    Evo 2 is an AI model that treats DNA as a language. It was built by scientists at the Arc Institute and NVIDIA with collaborators at Stanford University, UC Berkeley and UC San Francisco, and released with a preprint on 19 February 2025.[1] It was published in Nature on 4 March 2026.[2]

    Training and design

    The Nature paper describes a foundation model trained on 9 trillion DNA base pairs from a curated atlas spanning all domains of life. It has a 1-million-token context window at single-nucleotide resolution.[2] At release the Arc Institute put the training set at more than 9.3 trillion nucleotides from over 128,000 whole genomes plus metagenomic data.[4] It was trained for several months on more than 2,000 NVIDIA H100 GPUs.[5]

    Evo 2 uses the StripedHyena 2 architecture. The Arc Institute says this allowed training on 30 times more data than its predecessor, Evo 1, and reasoning over 8 times as many nucleotides at once.[6]

    What it can do

    The Nature paper reports that Evo 2 predicts the functional effects of genetic variation, from noncoding pathogenic mutations to clinically significant BRCA1 variants, without task-specific fine-tuning.[7] At release the Arc Institute said it was over 90% accurate at classifying BRCA1 variants as benign or potentially pathogenic.[8] The model can also generate mitochondrial, prokaryotic and eukaryotic sequences at genome scale.[9]

    Openness and safety

    The team released Evo 2’s training data, training and inference code, and model weights.[3] As of November 2026 the model comes in 1B, 7B, 20B and 40B parameter versions and can also be run through an NVIDIA-hosted API.[10] Citing ethics and safety risks, the developers left pathogens that infect humans and other complex organisms out of the base training data.[11]

    Google DeepMind’s AlphaGenome also reads up to 1 million DNA base pairs. It was trained on experimental gene-regulation data from public consortia such as ENCODE and GTEx.[12][13] For the wider field, see the synthetic biology guide.

    AlphaGenome

    AlphaGenome is Google DeepMind's AI model that predicts how DNA variants affect gene regulation. Inputs, accuracy, access, limits and 2026 updates.

    More in AI for science

    Developing in AI & Computing

    Go deeper

    Questions readers ask

    What is Evo 2?

    A biological foundation model that learns from DNA sequences across all domains of life. It can score genetic variants and generate new DNA sequences.[2][7][9]

    How much data was it trained on?

    The Nature paper says 9 trillion DNA base pairs. At release the Arc Institute described more than 9.3 trillion nucleotides from over 128,000 whole genomes plus metagenomic data.[2][4]

    Can Evo 2 tell harmful mutations from harmless ones?

    The Nature paper reports that it predicts the effects of genetic variation, including BRCA1 variants, without task-specific fine-tuning. The Arc Institute said it was over 90% accurate at classifying BRCA1 variants as benign or potentially pathogenic.[7][8]

    Is Evo 2 open source?

    Yes. The team released the training data, training and inference code, and model weights. It can also be run through an NVIDIA-hosted API.[3][10]

    Could Evo 2 be misused to design pathogens?

    Its developers left pathogens that infect humans and other complex organisms out of the base training data, citing ethics and safety risks.[11]

    Sources

    · 13 claims from 4 publishers

    Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.

    1. [1]

      The Arc Institute and NVIDIA released Evo 2, a genome language model, with a preprint posted on 19 February 2025. confirmedas of 2025-02-19

    2. [2]

      Evo 2 was published in Nature on 4 March 2026; the paper describes a foundation model trained on 9 trillion DNA base pairs from all domains of life, with a 1-million-token context window at single-nucleotide resolution. confirmedas of 2026-03-04

    3. [3]

      The Evo team released Evo 2's training data, training and inference code, and model weights. confirmedas of 2025-02-19

    4. [4]

      At release, the Arc Institute said Evo 2 was trained on more than 9.3 trillion nucleotides from over 128,000 whole genomes plus metagenomic data. confirmedas of 2025-02-19

    5. [5]

      Evo 2 was trained for several months on more than 2,000 NVIDIA H100 GPUs. confirmedas of 2025-02-19

    6. [6]

      Evo 2 uses the StripedHyena 2 architecture, which the Arc Institute says allowed training on 30 times more data than Evo 1 and reasoning over 8 times as many nucleotides at once. confirmedas of 2025-02-19

    7. [7]

      The Nature paper reports that Evo 2 predicts the functional effects of genetic variation, including noncoding pathogenic mutations and BRCA1 variants, without task-specific fine-tuning. confirmedas of 2026-03-04

    8. [8]

      The Arc Institute says Evo 2 was over 90% accurate in classifying BRCA1 variants as benign or potentially pathogenic. confirmedas of 2025-02-19

    9. [9]

      The Nature paper reports that Evo 2 generates mitochondrial, prokaryotic and eukaryotic sequences at genome scale. confirmedas of 2026-03-04

    10. [10]

      As of November 2026 Evo 2 is available in 1B, 7B, 20B and 40B parameter versions, and can also be run through an NVIDIA-hosted API. confirmedas of 2026-11-14

    11. [11]

      The developers left pathogens that infect humans and other complex organisms out of Evo 2's base training data, citing ethics and safety risks. confirmedas of 2025-02-19

    12. [12]

      AlphaGenome takes up to 1 million DNA base pairs as input, predicts thousands of molecular properties describing their regulatory activity, such as where genes start and end, splicing and RNA levels, and scores variants by comparing predictions for mutated and unmutated sequences. confirmedas of 2025-06-25

    13. [13]

      AlphaGenome was trained on data from public consortia including ENCODE, GTEx, 4D Nucleome and FANTOM5, covering hundreds of human and mouse cell types and tissues. confirmedas of 2025-06-25

    Revision history (1)
    1. Page created.

    Created Nov 14, 2026. Last reviewed by an editor on Nov 15, 2026. Next scheduled review: Nov 14, 2027.

    Cite this page

    "Evo 2." ContentLora, updated Nov 14, 2026. https://contentlora.com/wiki/evo-2

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.