Evo 2: DNA language model for genome modeling and design

Evo 2 is a foundation model for DNA sequence analysis that uses deep learning architectures to generate and score genomic sequences at scale. The model is available in multiple parameter sizes including a 7 billion parameter variant, and operates by learning probabilistic distributions over nucleotide sequences to enable both generative and discriminative tasks on DNA. It implements sampling strategies including temperature scaling, top-k filtering, and nucleus sampling to control the stochasticity and diversity of generated sequences. The model supports mixed-precision computation with bfloat16 tensors for memory efficiency and can be deployed on both GPU and CPU hardware depending on computational requirements. Generated outputs maintain biologically relevant properties such as GC content ratios while allowing for variant generation through configurable sampling parameters.