Dolphin is ByteDance’s open-source vision-language model for document image parsing: it converts scanned pages and PDF documents into structured, machine-readable output. Presented at ACL 2025, the official repository has collected more than 9,000 stars on GitHub and was still updated in 2026.
Document parsing is the unglamorous bottleneck of most RAG and digitization pipelines — tables, formulas, reading order and layout all break naive OCR. Dolphin approaches it as a two-stage vision-language task: it first analyzes the page layout, then parses each element in parallel using heterogeneous anchor prompts tailored to text, tables, formulas and figures. The result is a faithful structured representation of the page (markdown-style output) that downstream LLM systems can actually reason over, with strong accuracy on complex, dense documents. Because a single compact vision-language model does the whole job — rather than a chain of separate layout, OCR, table and formula tools — Dolphin is simpler to deploy and keeps global page context available while each element is interpreted, which reduces the reading-order and cross-element errors that plague stitched-together pipelines.
Dolphin is free to use from its public GitHub repository, including the released model weights; you supply your own inference compute.
Because a single compact vision-language model does the whole parsing job, Dolphin is simpler to operate than a chain of separate layout, OCR, table and formula tools, and its parallel element parsing keeps throughput high on long, multi-page documents.
It is for developers building document intelligence — RAG over PDFs, archive digitization, form and report processing — who need structured, layout-aware parsing instead of raw OCR text.