Unlimited OCR Works aims to revolutionize document parsing by enabling "one-shot long-horizon parsing," allowing for efficient and comprehensive information extraction from complex documents. This project targets researchers and developers in document analysis and AI, offering a powerful tool to process single images and multi-page documents, including PDFs, with advanced capabilities.

How It Works

This project builds upon Deepseek-OCR, utilizing Huggingface Transformers for inference on NVIDIA GPUs. It employs a transformer-based architecture with two configurations for single-image processing: 'gundam' (640px image size, crop mode) and 'base' (1024px image size, full image). For multi-page documents and PDFs, it uses the 'base' configuration, converting PDF pages into images via PyMuPDF for sequential processing. Key parameters like max_length=32768 and no_repeat_ngram_size=35 are configured to handle extensive output generation.

Quick Start & Requirements

• Primary Install/Run:
  • Huggingface Transformers: Requires Python 3.12.3, CUDA 12.9, torch==2.10.0, transformers==4.57.1, pymupdf==1.27.2.2.
  • SGLang: Install SGLang wheel, kernels==0.11.7, pymupdf==1.27.2.2. Launch server: python -m sglang.launch_server --model baidu/Unlimited-OCR --served-model-name Unlimited-OCR --attention-backend fa3 --context-length 32768.
• Prerequisites: NVIDIA GPU with CUDA 12.9.
• Links:
  • Paper: https://arxiv.org/abs/2606.23050
  • Model: baidu/Unlimited-OCR on Huggingface Transformers/ModelScope.

Highlighted Details

• Enables "one-shot long-horizon parsing" for complex document analysis.
• Offers distinct inference modes ('gundam' vs. 'base') for single images, balancing resolution and crop.
• Supports multi-page and PDF document processing through image conversion.
• Features extensive context length (max_length=32768) and advanced text generation controls.

Maintenance & Community

The project acknowledges support from the ModelScope community. Specific details regarding active contributors, community channels (e.g., Discord/Slack), or a public roadmap are not provided in the README excerpt.

Licensing & Compatibility

The README does not specify a software license. Therefore, licensing terms, commercial use permissions, and compatibility for closed-source integration remain undetermined.

Limitations & Caveats

Released in June 2026, this project appears to be recent. The README does not detail specific limitations, alpha status, known bugs, or unsupported platforms. The requirement for CUDA 12.9 indicates a dependency on recent NVIDIA driver and CUDA toolkit versions.