Dolphin
title: Dolphin (ByteDance) radar_quadrant: Languages & Frameworks radar_ring: Assess radar_position: inner created: 2026-05-26 last_updated: 2026-05-26 tags: [document-parsing, vision-model, llm, rag, ocr, pdf, bytedance, acl2025] source_url: https://github.com/ByteDance/Dolphin
A document image parsing model from ByteDance published at ACL 2025. Uses heterogeneous anchor prompting to parse complex document layouts — tables, figures, multi-column text, captions — from page images rather than treating documents as text streams. Model weights are open-sourced.
Conventional document parsers process a page sequentially, often scrambling multi-column layouts or losing table structure. Dolphin identifies structural anchors (table headers, figure captions, section boundaries) across different element types simultaneously, then parses content relative to those anchors. This preserves layout semantics: tables remain structured, figures retain their captions, and columns are not concatenated into a single stream.
The "heterogeneous" in heterogeneous anchor prompting refers to anchors drawn from different element types at the same time — not just text anchors, but visual and structural ones together.
Three open-source document parsers occupy distinct positions on this radar:
- Marker — fast, layout-aware pipeline; good for batches of PDFs with standard layouts
- docling — IBM production-grade; strong on table-heavy enterprise documents, LlamaIndex/LangChain integrations
- Dolphin — vision-model approach; strongest on complex mixed-layout documents (figures + tables + text interleaved); ACL peer review adds credibility
Dolphin's differentiator is the image-first, anchor-based approach. It handles document types where pipeline-based tools struggle: scanned documents with irregular layouts, academic papers with dense figure/table interleaving, and forms.
Feeds into RagChunkingStrategies and SearchBasedRAG as the ingestion layer — parse quality is a first-order determinant of RAG accuracy (confirmed by EnterpriseRAGLessons). Complements Markitdown (broad format conversion) and Docling (enterprise table-heavy docs) as a third ingestion option optimised for complex visual layouts. Langextract operates downstream of Dolphin: Dolphin converts images to structured text; langextract extracts typed fields from that text.
Placed in Languages & Frameworks / Assess / inner. ACL 2025 peer review and ByteDance open-source release make this a credible option for teams handling complex document layouts where Marker or docling underperform. Inner position reflects direct applicability to any RAG pipeline ingesting scanned documents, academic papers, or mixed-layout PDFs. Trial gate: batch parse of 20+ complex-layout documents with output quality compared against Marker or docling on the same corpus.
