OvisOCR3

OvisOCR3

Reads document images and returns text, structured JSON, Markdown with formulas and tables, key fields and figure captions from one 0.8B model. For developers who build OCR and document pipelines.

Made by ATH-MaaS 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: OvisOCR3 reads documents. It is a small model with 0.8B parameters that handles both pictures and text, and it starts from Qwen3.5-0.8B. One model covers document parsing, general OCR, key information extraction, figure captioning and cross-page table reconstruction. You pick the job by the prompt you give it, so you do not deploy a separate model for each task.

Who it suits: It suits developers and data teams who turn scanned pages, reports and forms into structured text and want one small model for the whole job.

Industry | Who uses it | How they use it | Example
Document processing | Data engineers | Convert page images into Markdown or structured JSON | Turning a scanned report into Markdown with tables
Back-office work | Operations teams | Pull named fields from documents into key-value records | Filling a JSON template from a page image
Publishing and research | Researchers | Capture formulas and figure captions from papers | Getting LaTeX for a formula on a page
Data migration | Archive teams | Rebuild tables that run across consecutive pages | Joining a split table into one HTML table

You can skip it if you need a chat assistant or a model for tasks other than reading documents.

Where to use it:

• Python projects that use the transformers library
• The online demo space that the card links for a quick test
• Windows, macOS or Linux with Python, to download the weights

Use case | Who | What they get back
Parse a page | Developers | A structured JSON file, a layout JSON file, or Markdown that carries formulas and tables
Extract fields | Analysts | Key-value records or JSON that follows your template
Caption a figure | Researchers | Localized captions, with Mermaid for flowcharts
Join split tables | Data teams | A complete HTML table with its source pages

How to set it up: You need Python and a machine that can hold the weights, about 1.7 GB in BF16.

• Python on Windows, macOS or Linux
• The huggingface_hub package
• The transformers library to run the model
• The task prompt files from the model repository

1. Install huggingface_hub.
2. Download the model files to your computer.
3. Read the task prompts in the prompts folder, since they define each output format.
4. Load the model with transformers and send a page image with the prompt for your task.

pip install -U huggingface_hub
hf download ATH-MaaS/OvisOCR3

How to use it: You choose the task by the prompt. The card says its prompt files set the output contract, and its own examples use those files directly. The visible card text gives no run command beyond the download.

hf download ATH-MaaS/OvisOCR3

Tool | What it does
Document parsing | Turns a page image into structured JSON, layout JSON, or Markdown
General OCR | Reads an image and gives back text, box positions, formulas, chart values, a class label or an answer
Key information extraction | Returns key-value records or JSON that follows your template
Document figure captioning | Returns page content with localized captions
Cross-page table reconstruction | Returns complete HTML tables from consecutive pages

Example: You send a page image together with the parsing-mode prompt. The model returns the page as Markdown with LaTeX formulas and HTML tables, or as structured JSON, depending on the mode you chose.

For newcomers and experts: If you are new to OCR, think of this as a reader that looks at a page image and writes down what is on it. A first small try is one page image with the parsing prompt, then reading the Markdown it returns.

For people who know the field, training ran in four steps: more pretraining, a middle phase, on-policy distillation, then reinforcement learning. In its own tables it leads most columns, though Xiaomi-OCR-0 scores slightly higher on OmniDocBench1.6, so check your own documents.

Title: OvisOCR3 document parsing scores
OmniDocBench1.6: 96.74
MDPBench: 88.80
PureDocBench: 82.31
Real5-OmniDocBench: 93.63

Page image and a task prompt go in
The 0.8B model reads and structures the page
Output is JSON, Markdown, HTML or text
You use the result in your pipeline

Good to know: The benchmark tables come from the model's authors and compare it against other OCR models of similar size.

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the page images you feed it and the downloaded model files
• Not for: chat or general text tasks, since the card describes OCR and document understanding

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: 203,064
• Likes: 121
• License: apache-2.0
• Based on: Qwen/Qwen3.5-0.8B
• Quantized versions: 7
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, document-parsing, multimodal, markdown, tables, formulas, key-information-extraction, figure-captioning, vllm

Links

• Model: https://huggingface.co/ATH-MaaS/OvisOCR3
• Files: https://huggingface.co/ATH-MaaS/OvisOCR3/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ATH-MaaS/OvisOCR3

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