Dolphin

Dolphin

Reads a document page image and extracts its text, tables, formulas and figures in reading order, in two stages. A compact 398M model for Chinese and English documents.

Made by ByteDance on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Dolphin is a document image parsing model from ByteDance. It looks at a page that mixes text paragraphs, figures, formulas and tables, and turns it into structured content. It works in two stages, analyzing the page layout first and then parsing each element.

Who it suits: It suits developers and teams who need to pull text, tables and formulas out of scanned or photographed documents. The record lists Chinese and English as its languages.

Title: Who uses it
Industry | Who uses it | How they use it | Example
Document processing | Backend developers | Convert page images into structured text | Parse a scanned report page by page
Research | Data teams | Extract formulas and tables from papers | Pull a table out of a paper page
Business operations | Automation builders | Read paragraphs and tables from forms and files | Parse a single table image

People who work with plain text files, with no page images, can skip it.

Where to use it:

• Python with the transformers library
• The authors' GitHub project, which holds the usage scripts
• Windows, macOS or Linux, to download the files

Title: Top jobs
Use case | Who | What they get back
Page-wise parsing | Developers | The elements of a whole page, in reading order
Element-wise parsing | Data teams | The content of one paragraph, table or formula image
Layout analysis | Researchers | A sequence of page elements in natural reading order

How to set it up:

• Python on your computer
• The huggingface_hub package
• The authors' GitHub repository, for the demo scripts

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Open the authors' repository and follow its usage instructions.

pip install -U huggingface_hub
hf download ByteDance/Dolphin

How to use it: The model card sends you to the authors' GitHub repository for detailed usage. It names two ways to run the model, and a natural language prompt controls the parsing task.

Title: Two ways to parse
Tool | What it does
Page-wise parsing | Handles an entire document image
Element-wise parsing | Handles one element image, such as a paragraph, table or formula

Example: You give the page-wise route an image of a document page that holds paragraphs and a table. The model first lays out the page elements in reading order, then parses each one. The card shows no sample output, so check the results against your page.

For newcomers and experts: Think of it as a reader for page images that understands layout. For a first try, download the files and run the page-wise route from the authors' repository on one clear page.

For experts, it is a vision-encoder-decoder model using a Swin Transformer encoder and an MBart text decoder. It is implemented as a VisionEncoderDecoderModel in Transformers. Stage two parses elements in parallel using heterogeneous anchors and task-specific prompts, which the authors credit for efficiency.

Page image goes in
Stage 1 lays out elements in reading order
Stage 2 parses elements in parallel
Structured content comes out

Good to know:

• Usage lives in the repo: the model card itself gives no run code, only a pointer.
• Languages: Chinese and English are listed.
• License: MIT.
• Cost to run: none named. The model is small, at 398M parameters.
• Data it reaches: the document images you give it, and downloads from Hugging Face.
• Not for: plain text with no page images.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 398M
• Weights: about 795 MB in F16; about 199 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 469
• Likes: 521
• License: mit
• Languages: zh, en
• Quantized versions: 2
• Fine-tunes: 3
• Tags: transformers, safetensors, vision-encoder-decoder, image-text-to-text, document-parsing, document-understanding, document-intelligence, layout-analysis, table-extraction, multimodal, vision-language-model

Links

• Model: https://huggingface.co/ByteDance/Dolphin
• Files: https://huggingface.co/ByteDance/Dolphin/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ByteDance/Dolphin

Then follow the usage in its model card. Nothing in this kit runs by itself.