Unlimited-OCR
Open Weights
Baidu
Unlimited-OCR is Baidu's open-weight document-parsing vision-language model, presented on 22 June 2026 under the MIT licence. It targets long-horizon parsing: one call can cover a multi-page image set or an entire PDF through an infer_multi path, alongside single-image parsing in either a tiled 'gundam' configuration or an untiled 'base' configuration. Parsed output carries <|det|> markers giving each block's element category and bounding box. Baidu ships the weights on Hugging Face and ModelScope, an official vLLM recipe with prebuilt Docker images, an SGLang server path, and a hosted document-parsing endpoint in the Baidu Cloud OCR product.
Strengths
- MIT-licensed open weights published on Baidu's own Hugging Face org and GitHub repository
- Multi-page image sets and whole PDFs parsed in a single infer_multi call, not one image at a time
- Two single-image configurations: tiled gundam (base_size 1024, image_size 640, crop_mode) and untiled base (image_size 1024)
- Emits <|det|> markers carrying element category and bounding box ahead of each parsed block
- Official vLLM recipe with prebuilt vllm-openai:unlimited-ocr images for CUDA 13.0 and a Hopper CUDA 12.9 variant
- SGLang launch path exposing an OpenAI-compatible chat-completions endpoint with streaming responses
- Hosted as a document-parsing (Unlimited-OCR) API in Baidu Cloud's OCR product, with a ModelScope mirror and ms-swift fine-tuning support
Caveats
- The model card and GitHub README publish no benchmark scores as text - only a demo GIF and an arXiv link - so no evaluation figures are recorded on this entry
- No parameter count and no advertised context window are stated by the vendor; the SGLang example merely launches the server with --context-length 32768, which is a serving flag rather than a published spec
- Inference requires trust_remote_code=True; the architecture is not supported by stock transformers
- Repetition is a known failure mode: every vendor example pins no_repeat_ngram_size=35 with an explicit n-gram window, and OmniDocBench evaluation additionally needs a vendor-supplied post-processing routine to strip <|det|> markers
- Multi-page and PDF parsing are restricted to base mode (image_size 1024); the higher-resolution gundam configuration works on single images only
- Vendor
- Baidu
- License
- MIT
- Release Date
- 2026-06-22
- Modalities
- textimage
Capabilities
Vision
Audio
Video
Tool Use