# Evo 2 (Arc Institute, NVIDIA)

Source: https://onco.cc/technologies/evo2/  
OnCo record `evo2` (Technology). Data CC BY-NC 4.0, attribute "Data from OnCo (onco.cc)"; commercial use needs a licence.

## TL;DR

A DNA language model trained on 9.3 trillion bases that can flag cancer-causing BRCA1 variants without being told about them.

## Summary

Evo 2 is a DNA language model from the Arc Institute and NVIDIA built on the StripedHyena architecture, which allows a context of 1M tokens so it can read long stretches of genome at once. The 2025 bioRxiv preprint describes 7B and 40B parameter models trained on 9.3 trillion bases, and shows zero-shot variant effect prediction, including classifying BRCA1 variants as pathogenic or benign without being trained on that task, alongside generation of genomic sequences. Weights are open. For oncology the interest is in interpreting variants in cancer genes and non-coding regions. Its human regulatory prediction is weaker than specialised models such as Enformer, so it complements rather than replaces them. For a newcomer: Evo 2 learned the language of DNA well enough to spot harmful BRCA1 mutations on its own.

## Fields

- Kind: Technology
- Status: emerging
- Last checked: 2026-09-08
- Tags: foundation-model; genome
- Principle: StripedHyena long-context sequence model over nucleotides.
- Since: 2025
- Strengths: Zero-shot variant effect; Open
- Limitations: Human regulatory prediction weaker than specialised models

## Sources

- Evo 2 bioRxiv 2025: https://www.biorxiv.org/content/10.1101/2025.02.18.638918v1

## Connected records

- fronts: [AI & Computation](https://onco.cc/fronts/ai-computation/)
- targets: [BRCA1 / BRCA2 (HRD)](https://onco.cc/targets/brca/)
- companies: [NVIDIA](https://onco.cc/companies/nvidia/)
- institutions: [Arc Institute](https://onco.cc/institutions/arc-institute/), [Stanford Health Care / Stanford Cancer Institute](https://onco.cc/institutions/stanford/)
- technologies: [NVIDIA BioNeMo](https://onco.cc/technologies/bionemo/)
- terms: [Autoregressive (next-token) modelling](https://onco.cc/terms/autoregressive-modelling/), [Genomic and protein language models: Evo 2, Enformer, ESM](https://onco.cc/terms/genomic-and-protein-language-models/), [Variant effect prediction](https://onco.cc/terms/variant-effect-prediction/)

---
JSON: https://onco.cc/api/v1/entities/evo2.json