Evo 2 opens genome-scale DNA AI

Evo 2 is the rare biology model release that feels like real platform infrastructure, not just a one-off paper result. The bigger question now is not whether the model is impressive, but how much of its zero-shot genomic intuition will survive contact with wet-lab reality.

• –The scale jump is serious: Arc says Evo 2 extends Evo 1 with 30x more training data and 8x longer context, which matters because genome function often depends on long-range structure
• –Open release is a big deal here because the team published model weights, training code, inference code, and the OpenGenome2 dataset instead of keeping the stack closed
• –The most developer-interesting result is generalization: one model handles bacteria, archaea, and eukaryotes well enough to spot regulatory DNA, splice sites, and pathogenic mutations without task-specific fine-tuning
• –The strongest practical use case is likely variant interpretation and genome annotation, where saving researchers rounds of brute-force experimental screening could be immediately valuable
• –The caveat is biological validation: sequence generation and regulatory design are exciting, but the gap between plausible DNA and reliably useful biology is still huge