🎯 Single Image OCR
Two processing modes: Gundam (base_size=1024, with cropping) for complex layouts, and Base (image_size=1024) for standard documents.
Coming Soon
Unlimited-OCR
Next-generation OCR engine by Baidu. One-shot long-horizon document parsing with a 32K context window — covering single images, multi-page documents, and PDFs. Coming soon to PDF to MD.
One-shot Parsing32K ContextPDF & ImageMIT LicenseSGLang Serving
32,768 Token context length
300 DPI PDF render precision
2 Modes Gundam / Base modes
MIT Open-source license
What is Unlimited-OCR
Unlimited-OCR is an open-source OCR system by Baidu, positioned as a next-generation document parsing solution beyond DeepSeek-OCR. It processes ultra-long documents in a single inference pass, achieving true One-shot Long-horizon Parsing.
📄 Multi-page Documents
Supports multi-image sequence input with automatic multi-page joint parsing. Anti-repeat mechanism avoids redundant output.
📑 Direct PDF Parsing
Renders PDF pages to 300 DPI images via PyMuPDF for multi-page parsing. No preprocessing needed.
⚡ High-Performance Inference
Dual backends: Hugging Face Transformers and SGLang. SGLang provides OpenAI-compatible API with FlashAttention 3.
🔧 Batch Processing
Built-in infer.py script auto-launches SGLang server with concurrent requests. Supports image directories or PDF input.
🏗️ Tech Stack
PyTorch with bfloat16 precision on NVIDIA GPU. Requires Python 3.12, CUDA 12.9, Transformers 4.57.1.
Coming to PDF to MD
We are integrating Unlimited-OCR into the PDF to MD platform, bringing more powerful document parsing capabilities. Once integrated, you can access all features through our API.
🌐 Online Experience
Upload documents directly on the PDF to MD website and select the Unlimited-OCR engine to experience one-shot long-document parsing.
🔌 Developer API
RESTful API coming soon — supporting single image, multi-page, and PDF calls with streaming, batch processing, and custom parameters.
📊 Structured Output
Parsed results output directly as Markdown, preserving tables, lists, and heading hierarchies for seamless LLM workflow integration.
🛡️ Stable & Reliable
Built on the SGLang high-performance inference framework with auto-scaling and fault recovery for enterprise-grade OCR service.
How It Works
From document upload to structured output — fully automated.
01 Upload PDF / Image
02 Unlimited-OCR Parsing
03 Structured Markdown
04 API / Download Output
Get Early Access
The Unlimited-OCR API is launching soon. If you would like early access, pricing details, or technical collaboration, please reach out.