By Arc InstituteLearning Objectives
- Understand what a genome language model is and how Evo 2 differs from a protein-structure model
- Explain what the August 2026 bacteriophage result actually demonstrated, and what it did not
- Evaluate the biosecurity safeguards the authors applied and where their limits lie
What Is Evo 2?
Evo 2 is a DNA language model from the Arc Institute, a non-profit research organization, developed with collaborators at Stanford. Where a large language model learns the statistical structure of text, Evo 2 learns the structure of genomes — predicting DNA one nucleotide at a time, across context windows reaching up to one million base pairs.
It was pretrained on OpenGenome2, a publicly released dataset of roughly 8.8 trillion tokens drawn from all domains of life, using the StripedHyena 2 architecture. Checkpoints ship at 1 billion, 7 billion, 20 billion, and 40 billion parameters; the 7 billion parameter model runs on ordinary GPUs, while the 40 billion parameter version needs multiple H100-class accelerators.
💡Key Concept
Genome model, not a protein model. AlphaFold predicts the three-dimensional shape a given protein folds into. Evo 2 works one level up and in the other direction — it generates the DNA sequence itself, including the regulatory and structural regions that are not genes at all. That is why it can produce a whole genome rather than a single molecule, and it is the reason the phage result was possible.
The August 2026 Result — Designing Working Viruses
On August 6, 2026, Science published work by a Stanford and Arc Institute team led by Brian Hie, with Samuel King as lead author, that used Evo 1 and Evo 2 to design complete bacteriophage genomes from scratch. Bacteriophages are viruses that infect bacteria only — they cannot infect human, animal, or plant cells.
The team fine-tuned the models on roughly 15,000 Microviridae genomes, using the small, well-studied phage phi X 174 as a design template. The models generated thousands of candidate genomes; the team filtered these to about 300 distinct designs, synthesized them, and tested them against E. coli.
| Stage | Count | Notes |
|---|---|---|
| Candidate genomes generated | Thousands | Filtered computationally for likely viability |
| Designs synthesized and tested | About 300 | Built in the laboratory |
| Viable phages | 16 | Successfully infected E. coli |
| Outperformed the natural template | 3 | Outcompeted wild phi X 174 head to head |
Roughly a 5 percent success rate, then — low in absolute terms, and the honest headline is not the hit rate but that the number is above zero at all. Several of the working phages shared less than 95 percent of their genome with any known phage, which by the usual convention makes them new species. A cocktail combining the designs rapidly overcame E. coli strains that had already evolved resistance to the natural phage.
✅Tip
Why this points at antibiotic resistance. Phage therapy — treating bacterial infections with viruses instead of antibiotics — has been pursued for a century with limited success, largely because finding the right phage means searching nature for one that happens to work, and bacteria evolve resistance to it. Generating candidates on demand, and generating a cocktail that resistance cannot easily route around, changes the economics of that search.
Biosecurity
The authors' central safeguard is in the training data: human viruses were excluded from pretraining, and the fine-tuning set was restricted to a single family of bacteria-infecting phages. The demonstration that such exclusions can prevent a model from generating human pathogens is itself one of the paper's claimed contributions.
⚠️Warning
The safeguard is the authors', not the field's. Independent commentators broadly welcomed the result and then made the same point: this work establishes that AI can design viable, non-natural viruses. The restraint shown here — excluding human viruses, restricting fine-tuning, staying within phages — is a choice these particular researchers made, and nothing compels the next group to make it. At least one reviewer also noted the exclusion may not be perfectly complete. Treat the biosecurity question as open.
Pricing
- 1B, 7B, 20B, and 40B checkpoints
- Contexts from 8K to 1 million base pairs
- Self-host on your own GPUs
- 8.8 trillion tokens on HuggingFace
- The pretraining corpus, published openly
The Apache 2.0 license was verified on the repository rather than inferred from coverage. It is genuinely permissive: commercial use is allowed, with no revenue threshold, no field-of-use restriction, and no separate agreement to negotiate. The README carries no acceptable-use policy — which is worth noticing in both directions, given the subject matter.
Strengths
- Genuinely open — Apache 2.0 weights at four sizes, plus the full pretraining dataset, published rather than gated
- Whole-genome scale — contexts up to a million base pairs let it generate complete genomes, not single genes or proteins
- Experimentally validated — the phage result was synthesized and tested in a laboratory and peer-reviewed in Science, not left as a computational claim
- Runs on ordinary hardware — the 7 billion parameter checkpoint does not require a specialized cluster
- A real therapeutic path — on-demand phage cocktails are a credible line of attack on antibiotic-resistant infection
Limitations and Considerations
- About a 5 percent hit rate — 16 viable phages from roughly 300 synthesized designs; this is a proof of principle, not a reliable design pipeline
- Small genomes only — phi X 174 is among the smallest genomes in biology. Larger phages, let alone bacteria, are a different order of difficulty
- No mechanistic understanding — the model produces sequences that work without explaining why, so researchers still cannot say what made the successful designs succeed
- Nothing near a patient — phage therapy faces regulatory, manufacturing, and safety hurdles that this result does not address
- Dual-use by construction — the safeguards are training-data choices made by these authors, and the same architecture trained without them would not carry them
Related Tools
- AlphaFold — DeepMind's protein-structure model, the closest precedent for an open scientific model with broad adoption
- NVIDIA BioNeMo — a framework for training and serving biological foundation models
Key Takeaways
- Evo 2 is an open genome language model from the Arc Institute that reads and writes DNA at single-nucleotide resolution, across contexts up to a million base pairs
- It was pretrained on OpenGenome2, roughly 8.8 trillion tokens spanning all domains of life, and ships at 1, 7, 20, and 40 billion parameters under the Apache 2.0 license
- In Science on August 6, 2026, a Stanford and Arc team used Evo 1 and Evo 2 to design complete bacteriophage genomes — 16 of about 300 synthesized designs worked, and 3 outcompeted the natural strain
- A cocktail of the designed phages broke through E. coli that had already evolved resistance, which is the concrete case for phage therapy against antibiotic-resistant infection
- Human viruses were excluded from pretraining as a deliberate safeguard — but that restraint is a choice these authors made, and the result establishes that AI can design viable, non-natural viruses
