OvisOCR2

OvisOCR2

Document parsing model that reads a page image and writes Markdown in reading order, covering text, formulas and tables. About 0.8B parameters, run through vLLM. For developers who digitize documents.

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: OvisOCR2 is a small model that parses whole document pages. You give it a page image, and it writes one Markdown document in reading order, with text, formulas, tables and picture regions. It was built by post-training Qwen3.5-0.8B and runs without a separate layout pipeline.

Who it suits: It suits developers and data teams who turn scanned or photographed pages into text they can search, edit or feed to other tools. You need Python and a machine that can run vLLM.

Industry | Who uses it | How they use it | Example
Document processing | Data engineers | Convert page images into Markdown for search and review | Turning a scanned report into editable text
Research | Academics | Parse papers that hold formulas and tables | Getting LaTeX formulas out of a paper page
Back office | Operations teams | Pull tables from page images into a usable form | Reading a table as HTML from a form page
Software | Application developers | Add page parsing to an app through a short Python class | Calling the parser on a batch of images

If you only need to read one page now and then, you can skip the local setup and try the online demo on Hugging Face.

Where to use it: You run it from Python on your own machine, with vLLM doing the inference.

• Windows, macOS or Linux with Python, to download the files
• A Python script that uses vLLM, where the card sets GPU memory use
• The online demo page on Hugging Face, for a quick try

Use case | Who | What they get back
Parse a page image | Developers | One Markdown string in natural reading order
Read a table | Analysts | The table written as an HTML table
Read a formula | Researchers | The formula written as LaTeX

How to set it up: You need Python, the vLLM package at version 0.22.1, and Pillow, plus a GPU for the settings the card uses. The model loads by its Hugging Face name.

• Python on Windows, macOS or Linux
• vLLM 0.22.1 and Pillow
• A GPU, since the card's code sets how much GPU memory vLLM may use

1. Install vLLM and Pillow with the command below.
2. Optionally download the model files ahead of time.
3. Copy the card's parser class and point it at the model name.

pip install "vllm==0.22.1" pillow

pip install -U huggingface_hub
hf download ATH-MaaS/OvisOCR2

How to use it: Build the parser once, then hand it a list of page images. It returns one Markdown string per page. The card's code uses greedy decoding and allows up to 16384 new tokens per page.

Tool | What it does
OvisOCR2Parser | Loads the model with vLLM and sets the fixed parsing instruction
parse | Runs a list of page images and returns Markdown strings
filter_imgtags | A flag on parse that controls the image tags in the result

from PIL import Image
from vllm import LLM, SamplingParams

Example: You pass in an image of a report page that holds a paragraph, a formula and a table. You get back Markdown with the paragraph as plain text, the formula as LaTeX, the table as HTML and the page's charts as image tags.

For newcomers and experts: OCR means reading text from a picture. This model goes further and rebuilds a page, including tables and formulas, as Markdown. Start with one clear page image and check the Markdown against the original.

For experts, it is Qwen3.5-0.8B post-trained with a data engine that mixes real and synthetic pages, and a multi-stage recipe of SFT, RL and OPD. Images are scaled between 448 by 448 and 2880 by 2880 pixels, and the instruction asks for text without translation. The trade-off is a small deployment footprint, with one model parsing each page in a single pass.

A page image and a fixed instruction go in
The 0.8B model reads text, formulas and tables
It writes Markdown in natural reading order
Charts become image tags with box coordinates
You get one Markdown string per page

Good to know: The authors report 96.58 overall on OmniDocBench v1.6 and 75.06 on the Avg3 score of PureDocBench. They say no end-to-end model had led the OmniDocBench board before, only pipelines. Others have published 27 quantized and 5 fine-tuned versions.

• License: Apache 2.0
• Cost to run: none named, with no key or paid service in the source
• Data it reaches: the page images you give it and the model files fetched from the hub
• Not for: translating or rewording text, since the instruction tells it to keep the original wording

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: 28,050
• Likes: 504
• License: apache-2.0
• Based on: Qwen/Qwen3.5-0.8B
• Quantized versions: 27
• Fine-tunes: 5
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, document-parsing, multimodal, markdown, tables, formulas, vllm, conversational

Links

• Model: https://huggingface.co/ATH-MaaS/OvisOCR2
• Files: https://huggingface.co/ATH-MaaS/OvisOCR2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ATH-MaaS/OvisOCR2

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