Dolphin

ByteDance built Dolphin around an analyze-then-parse strategy: a single vision language model first reads a page and produces its layout structure, then uses the detected elements as heterogeneous anchors to parse text, tables, and formulas in parallel rather than sequentially. Dolphin-v2 runs on a Qwen2.5-VL-3B backbone, recognizes 21 element types with attribute extraction, and exports structured JSON or Markdown, scoring 89.78 overall on the OmniDocBench document parsing benchmark with gains in text edit distance, formula parsing, and table parsing over earlier versions. The method was published at ACL 2025 (arXiv 2505.14059) and the weights are downloadable from Hugging Face at ByteDance/Dolphin-v2. Running it means cloning the repository, installing dependencies with pip, and pulling the checkpoint with the Hugging Face CLI, and inference expects an NVIDIA GPU. Licensing needs care: the README still shows an MIT badge, but the LICENSE file in the repository is the Qwen Research License Agreement, which permits non-commercial use only, so commercial pipelines should confirm terms first. The repository has around 9,000 stars.