Scientists from Arc Institute and NVIDIA have built Evo 2, the largest DNA language model ever created, capable of designing complete microbial genomes. Trained on 9.3 trillion nucleotides from over 128,000 genomes spanning all domains of life — bacteria, archaea, viruses, plants, and humans — the model can process sequences up to 1 million nucleotides at once and generate genome-length DNA that resembles real organisms.

Beyond genome design, Evo 2 identifies disease-causing mutations in human genes with over 90% accuracy, including clinically significant BRCA1 variants linked to breast cancer, without any task-specific fine-tuning. The model was trained on over 2,000 NVIDIA H100 GPUs and uses 30 times more data than its predecessor Evo 1.

"Evo 2 has a generalist understanding of the tree of life that is useful for a multitude of tasks, from predicting disease-causing mutations to designing potential code for artificial life," said Arc Institute co-founder Patrick Hsu. The research was published in Nature on March 4, and all code, training data, and model weights are publicly available.