A public Baidu GitHub repository has presented Unlimited-OCR, a document-parsing project described by its developers as an effort to support one-shot, long-horizon optical character recognition. The repository provides deployment and inference instructions rather than a consumer application, with workflows aimed at developers working with images and PDF documents.
The available setup guidance centers on inference with Hugging Face Transformers and Nvidia GPUs. The maintainers say their listed requirements were tested with Python 3.12.3 and CUDA 12.9. That qualification is important for prospective users: it identifies a tested environment without establishing that every other hardware or software combination will work in the same way.
Unlimited-OCR's documentation also points to an official vLLM recipe for another deployment path. It tells users to select Docker images according to their GPU platform, indicating that the project anticipates differences between accelerator environments. The repository does not reduce installation to a single universal configuration; instead, it presents several routes that developers can match to their infrastructure.
A further option uses SGLang. For that setup, the instructions call for a locally supplied SGLang wheel, a pinned kernels package and PyMuPDF, which is used to convert PDF pages into images. The repository then describes launching a server and sending requests to an OpenAI-compatible API. Streaming responses are supported in the documented request flow, allowing generated output to arrive incrementally.
The distinction between individual images and longer documents runs through the examples. One workflow sends an image for document parsing, while another handles multiple pages. For PDFs, pages are first converted to image files and then passed through the multi-page process. This arrangement makes the image conversion stage visible to developers instead of presenting PDF handling as an unexplained built-in step.
For larger workloads, the included batch-inference guidance says an `infer.py` script can start the SGLang server automatically and issue concurrent requests for either an image directory or a PDF. That offers a route from isolated tests to processing groups of inputs, while leaving deployment and resource choices with the operator. The repository additionally notes that output intended for OmniDocBench evaluation requires post-processing.
The project materials thank Deepseek-OCR, Deepseek-OCR-2 and PaddleOCR for models and ideas that informed the work. That acknowledgement places Unlimited-OCR in an existing line of OCR and document-parsing development rather than describing it as an isolated system.
The repository's materials focus on how to run the model and connect it to common inference frameworks. They do not, in the supplied evidence, provide enough information to independently establish accuracy, speed or cost advantages. The release is therefore best understood as a developer-facing model and implementation package with several documented inference options. Claims about comparative performance would require additional evaluation beyond the deployment instructions provided.
