#4444th of 83 in Open modelsFor science, health and research, science
evo2
DNA language model for scoring, embedding and generating sequences up to 1M bp
- Stars
- 4.3k
- License
- Apache-2.0
- Last commit
- Jun 2026
- Last release
- Feb 2026
- Language
- Jupyter Notebook
Missed the latest GitHub check; these facts may be older.
Overview
Evo 2 is a DNA language model trained on OpenGenome2 (8.8 trillion tokens across all domains of life) using the StripedHyena 2 architecture at single-nucleotide resolution. This repo runs inference through Vortex: scoring sequence likelihoods, extracting intermediate-layer embeddings and generating DNA from prompts, with checkpoints from 1B to 40B parameters on HuggingFace. Example notebooks cover BRCA1 variant effect prediction, exon classification and sparse autoencoder features.
Who it is for: Genomics researchers running DNA scoring, embedding and design on NVIDIA GPUs
Strengths
- Context up to 1M base pairs on the 7B, 20B and 40B checkpoints
- 7B models run in bfloat16 without Transformer Engine or FP8 hardware
- Notebooks for BRCA1 zero-shot variant scoring, exon classification, generation and SAE features
- Open training data (OpenGenome2) and training path via Savanna or BioNeMo
Weaknesses
- 20B, 40B and 1B models need FP8 via Transformer Engine on Hopper GPUs
- 40B model needs multiple H100 GPUs
- Linux only officially; WSL2 limited, requires CUDA 12.1+ and Flash Attention build
- Forward passes on very long sequences may be slow in Vortex
What it needs
- GPU required
- Docker
- Needs Vortex, PyTorch, Flash Attention 2.8.0.post2, Transformer Engine 2.3.0 (20B, 40B, 1B), CUDA 12.1+, cuDNN 9.3+
- Models: evo2_40b, evo2_20b, evo2_7b, evo2_40b_base, evo2_7b_base
Also in Open models
See all 83| Rank | Project | Score |
|---|---|---|
| 1 | ruviewWiFi CSI sensing on ESP32 for presence, breathing, heart rate and pose | 94 out of 100 |
| 2 | layaSingle-pass multilingual classifier returning typed decisions with calibrated odds | 88 out of 100 |
| 3 | rf-detrReal-time transformer models for object detection, segmentation and keypoints | 82 out of 100 |
| #4 | timesfmPretrained time-series foundation model for zero-shot forecasting from Google Research | 78 out of 100 |
| #5 | open_clipPyTorch library for training and running CLIP-style image-text models | 78 out of 100 |