Unlimited-OCR

Unlimited-OCR

Parses document images and multi-page PDFs into text with a 3.3B vision-language model from Baidu. For developers who add OCR and document parsing to their own tools.

Made by baidu 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: Unlimited-OCR is a 3.3B-parameter image-to-text model from Baidu for reading documents. You give it a page image, several page images or a PDF turned into images, and it parses them into text. The authors present it as a step beyond Deepseek-OCR, and it runs through the transformers library.

Who it suits: It suits developers and teams who turn scanned pages and multi-page documents into text inside their own code.

Industry | Who uses it | How they use it | Example
Finance and legal | Operations teams | Parse scanned multi-page documents into text | A stack of scanned contract pages read in one call
Publishing and archives | Archivists | Digitize page images | Old scans turned into searchable text
Software | Developers | Add document parsing to an app | A service that calls the model on each uploaded page
AI research | Researchers | Study long-document parsing | Compare its output with Deepseek-OCR on the same pages

If you have no NVIDIA GPU, the transformers example here will not run as written, so you may want to skip it.

Where to use it:

• Python with the transformers library on NVIDIA GPUs
• vLLM, which the project lists as supported for inference
• ms-swift, which the project lists as supported for training
• Baidu Cloud and ModelScope, where the project says the model is available
• A Hugging Face Spaces demo, if you want to try it without installing anything

Use case | Who | What they get back
Single page parsing | Developers | Text parsed from one image, saved to the output folder you name
Multi-page parsing | Document teams | Parsed text for several page images in one call
PDF parsing | Data teams | Parsed text for a PDF after its pages become images

How to set it up: You need these first:

• Python and an NVIDIA GPU, since the example moves the model to CUDA
• The package versions below, which the project tested with Python 3.12.3 and CUDA 12.9
• About 6.7 GB of disk for the BF16 weights

1. Install the pinned package versions with pip.
2. Download the model files with the hub tool.
3. Load the tokenizer and model with the Python code below.

torch==2.10.0
torchvision==0.25.0
transformers==4.57.1
Pillow==12.1.1
matplotlib==3.10.8
einops==0.8.2
addict==2.4.0
easydict==1.13
pymupdf==1.27.2.2
psutil==7.2.2

pip install -U huggingface_hub
hf download baidu/Unlimited-OCR

import os
import torch
from transformers import AutoModel, AutoTokenizer
model_name = 'baidu/Unlimited-OCR'
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModel.from_pretrained(
    model_name,
    trust_remote_code=True,
    use_safetensors=True,
    torch_dtype=torch.bfloat16,
)
model = model.eval().cuda()

How to use it: After loading, call infer on one image, or infer_multi on several page images. A single image can use the gundam setting or the base setting. Several pages and PDFs use the base size only.

model.infer(
    tokenizer,
    prompt='<image>document parsing.',
    image_file='your_image.jpg',
    output_path='your/output/dir',
    base_size=1024, image_size=640, crop_mode=True,
    max_length=32768,
    no_repeat_ngram_size=35, ngram_window=128,
    save_results=True,
)

Tool | What it does
model.infer | Parses one image with the gundam or the base setting
model.infer_multi | Parses several page images in one call at image size 1024
PDF route | Turns each PDF page into an image with PyMuPDF at 300 dpi, then calls infer_multi

Example: You pass three page images, page1.png to page3.png, to infer_multi with the multi-page prompt. The model parses them together and saves the results to the output folder you set.

For newcomers and experts: This is an OCR model: it reads text out of pictures of pages. A first small thing to try is the Hugging Face Spaces demo, then the single-image example once the packages are installed.

Under the hood, gundam turns crop mode on at image size 640, while base turns it off at 1024. The examples set a maximum length of 32768 and a no-repeat n-gram size of 35, with a wider window for multi-page runs. The weights are BF16, and the 4-bit size is an estimate.

Page images or PDF pages go in
A prompt tells the model what to parse
The model reads the pages on the GPU
Parsed results save to your output folder

Good to know: Weight size depends on precision, and the 4-bit figure is an estimate.

Title: Weight size by precision
BF16: 6.7 GB
4-bit estimate: 1.7 GB

• Remote code: the example sets trust_remote_code, so it runs code that ships with the model
• Hardware: the transformers example targets NVIDIA GPUs
• Other versions: 40 quantized and 16 fine-tuned versions made by other people are listed
• License: MIT
• Cost to run: none named, you supply the GPU
• Data it reaches: the images and PDFs you point it at, and the output folder it writes to

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 3.3B
• Weights: about 6.7 GB in BF16; about 1.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 5,833,292
• Likes: 5,004
• License: mit
• Quantized versions: 40
• Fine-tunes: 16
• Tags: transformers, safetensors, unlimited_ocr, feature-extraction, baidu, vision-language, image-text-to-text, conversational, multilingual

Links

• Model: https://huggingface.co/baidu/Unlimited-OCR
• Files: https://huggingface.co/baidu/Unlimited-OCR/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download baidu/Unlimited-OCR

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