Arc Institute

weights

open(Apache-2.0 ungated checkpoint)

data

open(OpenGenome2 raw FASTA and augmented JSONL released)

code

open(Savanna training framework and Evo 2 training guidance released)

license

Apache-2.0(OSI)

Evo 2 releases all three parts of the from-scratch stack: permissive weights, the OpenGenome2 training corpus, and the named open training framework.

• https://huggingface.co/api/models/arcinstitute/evo2_7b recorded 2026-09-10

  license: apache-2.0; gated: false; private: false; evo2_7b.pt present.

• https://raw.githubusercontent.com/ArcInstitute/evo2/main/README.md recorded 2026-09-10

  Links downloadable checkpoints, the OpenGenome2 raw and processed training data, and the open Savanna framework under its Training and Finetuning section.

Downloads of the declared 7B checkpoint on Hugging Face are the adoption signal here, one size among Evo 2's several published checkpoints.

• https://huggingface.co/api/models/arcinstitute/evo2_7b recorded 2026-09-10

  downloads: 1350; gated: false; private: false.

Evo 2 documents two readable capabilities: an explicit training and fine-tuning pathway, and genuine DNA sequence generation rather than only prediction. Its context window scales from 8K to one million bases, but that is a longer window at the same single-nucleotide granularity throughout, not a second scale, and the range of model sizes offered does not count as an added capability either.

• https://raw.githubusercontent.com/ArcInstitute/evo2/main/README.md recorded 2026-09-10

  Documents training and fine-tuning, prompted DNA generation, and checkpoints spanning 8K, 262K, and one-million-base contexts.