Evo 1

Evo 1 is a 7-billion-parameter biological foundation model developed by researchers at the Arc Institute together with Stanford University, UC Berkeley, and Together AI, first released on February 27, 2024. It treats DNA as a language: trained on OpenGenome, a dataset of roughly 300 billion tokens drawn from whole prokaryotic genomes, it learns biological complexity from individual nucleotides up to whole genomes.

Built on the StripedHyena hybrid architecture, Evo 1 models sequences with near-linear scaling and a context length of up to 131 kilobases, enabling both prediction tasks and generative design from the molecular scale to the whole-genome scale. Its weights are released under the Apache 2.0 license, and the team also published intermediate pretraining checkpoints for open science.

The model family was later extended with Evo 2, and a 2026 Science paper showed Evo-designed bacteriophage genomes outperforming natural ones in some respects, a result flagged as a significant biosafety consideration. Catalogued here as the original 2024 release; Evo 2 is a separate entry.