Skip to content
Get packs

Unlimited OCR

Open-source OCR model from Baidu that parses long, multi-page documents into text.

GitHub Download source
Stars
30.3k
Forks
3.2k
Licence
MIT
Language
Python
Last update
4 days ago
Type
AI model / agent

What it is

Unlimited OCR is a document-parsing model from Baidu that reads images of pages and converts them into structured text. The README says its second version, Unlimited OCR 2, is designed to handle very long documents, including multi-page and hundred-page files, by using a sparse attention method called SR-SWA. It is released under the MIT license and can be run through Hugging Face Transformers or SGLang.

  • Parse a multi-page document in one request using the infer_multi method so pages share context
  • Parse a single page image with the infer method using dynamic tiling
  • Run an SGLang server to serve the model for multiple requests
  • Convert PDFs to page images with PyMuPDF before feeding them to the model

Great for

  • Developers who want to parse multi-page scanned documents with a locally hosted model
  • Teams that need to process long PDFs converted to page images
  • Researchers comparing long-document OCR models using the MarathonDocBench benchmark
  • Engineers who already run GPU-based model serving with SGLang

Not ideal for

  • People who need a simple point-and-click tool, since the README only shows Python and server commands
  • Users without an NVIDIA GPU, because the README's examples and tested setup use CUDA hardware (A800-80GB)
  • Anyone needing guaranteed accuracy on their own documents, since the README's benchmark claims are not independently verified here
Step-by-step guide: how to use Unlimited OCRInstall, first run, examples, tips and FAQs
DeepSeek-OCR vs Unlimited OCR

How to get it

  1. 1

    Install with a package manager

    You'll need: Python 3.10 · NVIDIA GPU with CUDA · uv (for the SGLang setup)

    Install Transformers dependencies

    pip install torch==2.13.0 torchvision pillow transformers==4.52.1 huggingface-hub==0.36.0 tokenizers==0.21.4 addict==2.4.0 einops==0.8.2 easydict==1.13

    Create SGLang virtualenv

    uv venv --python 3.10

    Install PyMuPDF

    uv pip install pymupdf==1.27.2.2
  2. 2

    Run it

    • Install the Transformers dependencies listed in the README.
    • Load the model with AutoModel.from_pretrained using baidu/Unlimited-OCR-2 and trust_remote_code=True.
    • Pass a list of page image files to model.infer_multi, in page order.
    • Read the returned text and check the output directory set by output_path.
  3. 3

    Get the most out of it

    • Use infer_multi for multi-page documents so the model can use cross-page context
    • Use the SGLang server for serving and tune the page-window environment variables for very long documents
    • Lower SGLang's SAM attention chunk cap to 64 to reduce peak memory, as the README suggests
    • Set PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to reduce memory fragmentation

Downloads

Source code (.zip)main branch · straight from GitHub

Strengths

  • Built to handle long documents in one pass, according to the README
  • Released under the MIT license with public model weights on Hugging Face
  • Offers both a simple Transformers path for quick tests and an SGLang path for serving

Trade-offs

  • Setup requires a GPU and specific pinned package versions, which can be fiddly
  • Performance claims come from the authors' own benchmarks and leaderboard images
  • Documentation is partly in code comments and the README is truncated, so some settings are unexplained

How it compares

ProjectStarsTypeLevelUpdatedBest for
Unlimited-OCR (this)30.3kAI model / agentadvanced4 days agoDevelopers who want to parse multi-page scanned documents with a locally hosted model
DeepSeek-OCR24.1kAI model / agentadvanced9 months agoDevelopers who want an open-weight OCR model they can run themselves

Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.