LightOnOCR-3-4B

LightOnOCR-3-4B

Reads document page images and returns the text as markdown, or labeled blocks with bounding boxes, image descriptions and chart data. For developers who build OCR and document-understanding pipelines.

Made by lightonai 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: LightOnOCR-3-4B is the largest model in LightOn's LightOnOCR-3 family of compact end-to-end OCR models. It reads a page image and writes out its text, and it can also return where each block sits, short image descriptions and the data behind charts. The authors recommend it for most OCR work.

Who it suits: It suits developers and data teams who convert scans, receipts, forms and reports into text or structured data and want a single model rather than a multi-step pipeline.

Industry | Who uses it | How they use it | Example
Document processing | Data engineers | Turn page images into markdown text | Converting a scanned report to markdown
Finance and admin | Back-office teams | Read receipts, forms and tables | Extracting a table from a form
Search and retrieval | Pipeline builders | Index text, image descriptions and chart data together | Making a chart searchable as a table
Research | Researchers | Test layout-aware OCR with bounding boxes | Checking block labels against a page

You can skip it if you need a tiny model, since the family also has 1B and 0.8B versions for speed.

Where to use it:

• Python with the transformers library
• OCR and document pipelines that take page images
• The online demo space that the card links, for a quick test
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Full-page transcription | Developers | The page text as markdown
Layout-aware extraction | Pipeline builders | Blocks with a label and a bounding box each
Chart reading | Analysts | An HTML table holding a chart's data points

How to set it up: You need Python, the download tool and a machine that can hold the weights, about 9.1 GB in BF16.

• Python on Windows, macOS or Linux
• The huggingface_hub package
• The transformers library to run the model
• A GPU or other hardware with enough memory

1. Install huggingface_hub.
2. Download the model files to your computer.
3. Load the model with transformers and send it a page image.

pip install -U huggingface_hub
hf download lightonai/LightOnOCR-3-4B

How to use it: The visible card names two modes and gives no run command. You call the model with an empty prompt for plain transcription, or with the grounding prompt for labeled blocks. Positions are given in page coordinates scaled from 0 to 1000.

Tool | What it does
Transcription mode | Takes an empty prompt and returns the page text as markdown
Grounding mode | Takes the grounding prompt and returns each block with a type label and a bounding box

Grounding output marks each block with its type and box, then gives the content that follows:

![title](80,70,920,140)
# Company report 2025

![text](80,175,920,225)
Revenue rose 20%.

Example: You send a report page with the grounding prompt. A heading block comes back labeled title with its box, a paragraph comes back labeled text, and a chart comes back as an HTML table of its data points.

For newcomers and experts: If you are new to OCR, it is software that reads text from a picture of a page. A first small try is one page image with an empty prompt, then comparing the markdown with the page.

For people who know the field, the 4B and 0.8B builds run on the Qwen3.5 vision-language design. The authors say that eases plugging into existing tools and gives a real speed-up. The 1B keeps the LightOnOCR-2 architecture and works as a drop-in upgrade.

A page image goes in
The prompt picks transcription or grounding
The model reads the page end to end
You get markdown or labeled blocks with boxes

Good to know: Labels in grounding mode include text, title, list, header, footer, page_number, footnote, caption and formula, and the card lists more.

• License: Apache 2.0, for research and commercial use
• Cost to run: none named
• Data it reaches: the page images you send and the downloaded model files
• Not for: pipelines that need a tiny footprint, since the family has 1B and 0.8B versions for that

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 4.5B
• Weights: about 9.1 GB in BF16; about 2.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 15,491
• Likes: 143
• License: apache-2.0
• Quantized versions: 16
• Fine-tunes: 4
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, document-understanding, visual-grounding, vision-language, tables, forms, conversational

Links

• Model: https://huggingface.co/lightonai/LightOnOCR-3-4B
• Files: https://huggingface.co/lightonai/LightOnOCR-3-4B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download lightonai/LightOnOCR-3-4B

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