Unlimited-OCR
A document-parsing OCR model from Baidu built for one-shot, long-horizon parsing, with inference support through Transformers, vLLM and SGLang.
Unlimited-OCR is an OCR model from Baidu, described as pushing DeepSeek-OCR further, aimed at what the project calls one-shot, long-horizon parsing. It can be run for inference using Hugging Face Transformers on NVIDIA GPUs, through vLLM using published Docker images for different GPU platforms, including a Hopper-specific build, or through SGLang, which exposes an OpenAI-compatible streaming API once its server is started.
The project also provides a batch inference script that starts an SGLang server automatically and sends concurrent requests for a directory of images or a PDF, letting a whole document set be processed as OCR in a single run. Since its release, the model has gained community support for training with ms-swift, hosting on Baidu Cloud, and a public demo on Hugging Face Spaces, and a paper describing the model is available on arXiv.