Evo 2

Evo 2 is a DNA foundation model from the Arc Institute trained across the tree of life, operating on genomic sequence at a scale that lets it reason about regulatory context rather than isolated genes. It is one of the most significant open releases in computational genomics, and the weights are genuinely available.

What it does

Trained on trillions of nucleotides spanning prokaryotic and eukaryotic genomes, Evo 2 handles long context — enough to capture regulatory elements and their targets in the same window. It supports variant effect prediction without task-specific training, and can generate biologically plausible sequence, which is where both the promise and the biosecurity conversation live.

Strengths

• Openly released weights, code and training data.
• Long genomic context rather than short-window models.
• Zero-shot variant effect prediction across species.
• Generative capability for synthetic sequence design.

Limitations

Running it meaningfully requires substantial GPU resources. Generative DNA models raise real biosecurity questions that the field is still working through, and the released model reflects deliberate choices about that. As with any foundation model, zero-shot performance varies by task and needs benchmarking against your specific question.