Unlimited OCR is an optical character recognition tool that performs one-shot long-horizon document parsing on images of arbitrary length.
The tool addresses the challenge of extracting text from documents that exceed typical model context windows or contain complex layouts. It uses a vision-language model approach that can process entire documents in a single pass without requiring multiple crops or sequential processing steps. The model is designed to handle documents of varying lengths and complexities through an extended context capability.
Teams should consider this tool if they need to extract text from full-page or multi-page document images without preprocessing them into smaller sections. It suits projects involving document digitization, form processing, or archival work where maintaining document context and layout understanding is important. The tool is available through multiple inference frameworks including Hugging Face Transformers, vLLM, and SGLang, allowing integration into different deployment environments. It also supports fine-tuning through the ms-swift framework for teams wanting to adapt it to domain-specific documents.
The project maintains active integration with major inference and training ecosystems, having added support for vLLM inference, SGLang deployment, and ms-swift training. The model is distributed across multiple platforms including Hugging Face, ModelScope, and Baidu Cloud, indicating broad accessibility. The tool provides Docker images for different GPU architectures and includes batch inference capabilities for processing image directories or PDFs concurrently.
