evo2

name	evo2
description	Evo 2 — long-context DNA language model for genome modeling and design across all domains of life (Arc Institute; StripedHyena 2 / OpenGenome2). Use when: (1) Autoregressive DNA likelihoods or embeddings at up to ~1 Mb context, (2) Sequence generation / design, (3) Contrasting generative gLMs with Nucleotide Transformer or gLM2. Upstream: https://github.com/ArcInstitute/evo2. Predecessor: Evo (Science 2024). Route via genome-language-model. Heavy GPU requirements.
license	MIT
category	analysis-tools
tags	["Evo2","Evo","gLM","DNA","generative","OpenGenome2","Arc-Institute"]
upstream	https://github.com/ArcInstitute/evo2
stage	function

Evo 2

Upstream: ArcInstitute/evo2 · Dataset: https://huggingface.co/datasets/arcinstitute/opengenome2

Citation

Brixi, G. et al. Genome modelling and design across all domains of life with Evo 2. Nature https://doi.org/10.1038/s41586-026-10176-5 (2026).

Nguyen, E. et al. Sequence modeling and design from molecular to genome scale with Evo. Science 386, eado9336 (2024). https://doi.org/10.1126/science.ado9336

See also docs/references.md.

Analytical thinking

Evo 2 is a generative / long-context DNA LM (single-nucleotide resolution, contexts up to ~1M bp) pretrained on OpenGenome2 (~8.8T tokens, all domains). Checkpoints include evo2_7b (widest hardware reach) through evo2_40b (FP8 / Hopper-class for full accuracy per upstream).

vs	Prefer Evo 2 when
nucleotide-transformer	Generation / long AR likelihoods across domains
omg / gLM2	Pure DNA generative modeling vs mixed AA+DNA metagenomic LM
Classical assemblers	Design/scoring sequences — not contig assembly

Intermediate-layer embeddings often work better than final layer (per paper). Validate outputs after hardware/config changes (upstream tests).

How to run

# light path: 7B without Transformer Engine — see upstream Installation
pip install evo2   # after system CUDA / deps per README
python - <<'PY'
from evo2 import Evo2
import torch
evo2_model = Evo2('evo2_7b')
sequence = 'ACGT'
input_ids = torch.tensor(
    evo2_model.tokenizer.tokenize(sequence), dtype=torch.int
).unsqueeze(0).to('cuda:0')
# forward / embeddings / generation — see upstream Usage sections
PY

Also: NVIDIA hosted API / NIM options documented upstream.

Decision tree

Long DNA LM / design?
├─ Generative + all-domains pretrain → evo2
├─ Multi-species tracks / NTv3 representation → nucleotide-transformer
├─ Efficient / long-range representation → dnabert2 / caduceus
├─ Metagenomic gLM2 / OMG corpus → omg
└─ Classical search → homology-search

Related skills

genome-language-model · nucleotide-transformer · dnabert2 · caduceus · omg · metagenomics-llm · lucaphylo · tool-selection