Evolla

An 80B-scale multimodal protein-language generative/chat model from Fajie Yuan's lab, pairing a PLM encoder with an LLM decoder via a compression-alignment module. Trained on 546M protein-QA pairs (~150B tokens) with RAG and DPO, and evaluated with a purpose-built IRS framework. Evolla was merged into HuggingFace Transformers (July 2025), and the released Evolla-10B checkpoint has ~81K HuggingFace downloads — strong real-world adoption for a science FM.