Baidu Releases Open-Source Unlimited OCR for Long-Horizon Document Parsing
The Chinese tech giant has made its latest optical character recognition model available on GitHub, citing contributions from Deepseek and PaddleOCR projects.
Baidu has released Unlimited OCR, an open-source software tool designed for one-shot long-horizon parsing. The project, hosted on GitHub under the baidu repository, aims to streamline the extraction of text from extensive documents in a single inference pass. The release positions the tool as a solution for handling complex document structures that typically require multi-stage processing in legacy systems.
The architecture relies on Hugging Face transformers for inference, with support for NVIDIA GPUs. According to the technical specifications provided by the developer, the environment requires Python 3.12.3 and CUDA 12.9. The project is managed via a uv-managed virtual environment, ensuring consistent dependency handling for developers integrating the tool into their workflows.
To facilitate integration, Unlimited OCR features an OpenAI-compatible API that supports streaming requests. This design allows developers to process data incrementally, which can be critical for managing memory usage when dealing with large-scale document sets. The tool also utilises PyMuPDF for converting PDF files into images prior to processing, ensuring compatibility with a wide range of document formats.
For high-volume operations, the system includes batch inference capabilities via the SGLang server. The infer.py script can automatically start the SGLang server and send concurrent requests to an image directory or PDF folder. This feature is intended to improve throughput for institutions processing large archives or requiring rapid turnaround on document digitisation tasks.
The release acknowledges contributions from several existing open-source projects, specifically Deepseek-OCR, Deepseek-OCR-2, and PaddleOCR. Baidu cited the valuable models and ideas from these initiatives as foundational to the development of Unlimited OCR. The project currently lacks detailed public performance metrics or specific accuracy rates for its long-horizon parsing capabilities, leaving the exact scope of its efficiency to be determined by independent testing.
