Evo 2

Evo 2 is a genome foundation model from the Arc Institute, built with NVIDIA and collaborators at Stanford, UC Berkeley, and UCSF — first released as a preprint in February 2025 and later published in Nature. It is a StripedHyena 2 model trained on 9.3 trillion nucleotides from over 128,000 whole genomes spanning bacteria, archaea, and eukaryotes including humans. The flagship is 40 billion parameters with a 1-million-nucleotide context window; 7B, 20B, and 1B variants cover different hardware realities. The team released everything: weights, training code, inference code, and the 8.8-trillion-token OpenGenome2 dataset — making it one of the most fully open scientific foundation model releases in any domain. It predicts disease-relevant mutation effects (over 90% accuracy distinguishing benign from pathogenic BRCA1 variants) and can design novel bacterial-scale DNA sequences.\n\nThe honest picture: pathogen data for human and animal pathogens was deliberately excluded from training, so it is not a bioweapon cookbook — but its ability to design phage genomes raised legitimate safety debate in 2026. The 40B model needs multiple Hopper GPUs; the 7B is the practical entry point. And genomic models have limits: generating novel genomes remains harder to evaluate than classifying known mutations.