LightOnOCR-3-0.8B

LightOnOCR-3-0.8B

A small 0.8B OCR model from LightOn that reads a page image into markdown, and in grounding mode also labels each block with a bounding box. The fast, light size of the LightOnOCR-3 family.

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-3-0.8B is the smallest model in LightOn's LightOnOCR-3 family of OCR models. It reads an image of a page and writes out the text, and it can also say where each block sits on the page.

It is one model rather than a multi-step document pipeline. The card says it copes with receipts and forms, tables and charts, pages laid out in columns, and mathematical formulas. The 0.8B size follows the Qwen3.5 vision-language design, which the card says eases integration and speeds it up.

Who it suits: Developers who need fast OCR and page understanding in their own apps, and teams who digitize documents and want structured output without a large model.

Industry | Who uses it | How they use it | Example
Finance | Back-office teams | Read receipts and forms | Turning scanned receipts into text
Legal and admin | Records teams | Digitize multi-column documents | Converting printed reports to markdown
Software | App developers | Add page understanding to a product | Extracting tables from uploaded pages
Research | Data teams | Pull chart data into tables | Turning a chart image into a data table

You can skip it if you need the highest accuracy, since the card names the 4B model as the best for most tasks, or if you only process plain text files.

Where to use it: It loads through the transformers library in Python, and the family also offers 1B and 4B sizes.

• Python with transformers
• Windows, macOS and Linux
• Pipelines that feed page images in and take markdown out

Use case | Who | What they get back
Page transcription | Developers | The page text as markdown
Layout grounding | Data teams | Each block with a label and bounding box
Chart to table | Analysts | An HTML table of the chart's data points
Image captions | Retrieval teams | A short description of each picture

How to set it up: You need Python and the Hugging Face tools. The model is small, at about 1.7 GB in BF16, so most modern machines can hold it.

• Python with pip and transformers
• About 1.7 GB of disk space
• A GPU is helpful for speed

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Load the model with transformers and pass it a page image.

pip install -U huggingface_hub
hf download lightonai/LightOnOCR-3-0.8B

How to use it: Pass a page image with an empty prompt to get the page text as markdown, the default mode and the same as LightOnOCR-2. Pass the word grounding as the prompt to get labels and boxes too.

Mode | What you get
Empty prompt | All text on the page as markdown
grounding prompt | Each block with a label and a bounding box, plus image descriptions and chart tables

In grounding mode each block starts with a placeholder showing its type and box, in page coordinates scaled from 0 to 1000. The card's example looks like this.

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

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

![image](80,285,460,545)
A solar-powered factory.

![chart](540,285,920,545)
<table>
  <tr><th>Year</th><th>Revenue</th></tr>
  <tr><td>2024</td><td>€10M</td></tr>
  <tr><td>2025</td><td>€12M</td></tr>
</table>

Example: You send an image of a company report page with the grounding prompt. You get a title block, a text block saying revenue rose 20 percent, an image block with a short description and a chart block as a table.

For newcomers and experts: If you are new, start with the empty prompt on one clear page image and read the markdown it returns. Try the grounding prompt next to see labels and boxes.

If you know the field, the block labels include text, title, list, header, footer, page number, footnote, caption, formula, code and table, plus image and chart blocks. Boxes use coordinates normalized to 0 to 1000, so they are independent of image size.

Page image in
Empty or grounding prompt
Model reads layout and text
Markdown out, with labels and boxes in grounding mode

Good to know: The family has three sizes, and the 0.8B is the fast, efficient one. The card says all models are released under Apache 2.0 for research and commercial use.

• License: Apache 2.0, for research and commercial use
• Cost to run: none named
• Data it reaches: the page images you give it, and the model files you download
• Not for: the highest accuracy, since the card recommends the 4B model for most tasks

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 853M
• Weights: about 1.7 GB in BF16; about 426 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 13,965
• Likes: 99
• License: apache-2.0
• Languages: en
• Quantized versions: 17
• 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-3-0.8B
• Files: https://huggingface.co/lightonai/LightOnOCR-3-0.8B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download lightonai/LightOnOCR-3-0.8B

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