Unlimited-OCR

Unlimited-OCR is an OCR model for developers building document extraction workflows on their own hardware or servers. It handles individual document images and multi-page inputs, including PDFs converted to page images.

• Document parsing: Process a single image or multiple pages together to extract document content.
• Batch processing: Run concurrent extraction jobs across an image directory or the pages of a PDF.
• API integration: Serve the model through SGLang with an OpenAI-compatible API that streams results to connected applications.
• Model training: Use ms-swift to train the model for your own document tasks.

The Python project is open source under the MIT license. Local inference supports Hugging Face Transformers on NVIDIA GPUs, with vLLM and SGLang as serving alternatives. vLLM Docker images include an option for NVIDIA Hopper GPUs. When you run inference locally, document processing takes place on your machine or server.

Baidu Cloud provides a separate hosted option, and Hugging Face Spaces has a browser demo. Those services run outside your local deployment. PDF processing uses page images rather than feeding the original PDF directly into the model.