Based on the linked paper, it reads like they took an existing open weight LLM, tacked on a scheme to parse and output genomics data, and finetuned it a bit. It’s the same approach chatgpt and other big-name models use for reading and producing images, but applied to a different data format.