Stanford University · v2 · 3× · naposledy 08. 8. 2026
Evo 2 is a genomic AI Foundation Model developed by Arc Institute in collaboration with NVIDIA, with researchers from Stanford University, UC Berkeley, and UCSF contributing to the project. It was trained on over 9.3 trillion nucleotides from more than 128,000 genomes across over 100,000 species from all three domains of life and can both analyze genetic sequences (such as identifying disease-causing mutations) and generate new DNA sequences up to the size of simple bacterial genomes. The model is based on the StripedHyena-2 architecture, processes contexts of up to 1 Megabase, and is available in variants from 1B to 40B parameters; code, training data, and weights are openly accessible (Apache-2.0 License) and integrated into NVIDIA's BioNeMo Framework.
Vlastnosti
| Compliance/Certification | No formal certification; ethical safeguards: exclusion of human-pathogen sequences from training data, oversight by Stanford bioethicist; NIM governed by NVIDIA Software License & Open Model License |
| Deployment Model | Local inference via GitHub/Docker/Apptainer, hosted NVIDIA API, self-hosted via NVIDIA NIM (Docker/Kubernetes/Cloud), cloud marketplaces (AWS SageMaker) |
| Use Case Scope | Genomics research, disease-causing mutation prediction, synthetic genome design, drug discovery, agriculture, biotechnology, materials science |
| Integrations | NVIDIA BioNeMo framework, NVIDIA NIM microservices, Amazon SageMaker JumpStart, AWS/GCP/Azure/DGX Cloud, Evo Designer, Goodfire interpretability visualizer |
| License | Apache License 2.0 (code/model); NIM container additionally under NVIDIA Open Model License Agreement |
| Platform | GitHub (ArcInstitute/evo2), NVIDIA BioNeMo/NIM, PyPI package, Hugging Face (OpenGenome2 dataset) |
| Price | Free / open source (weights, code, training data freely available); NVIDIA BioNeMo models are free to use |
| Release Date | February 19, 2025 (preprint & public release), Nature publication March 2026 |