Baidu’s OCR model treats page limits as a suggestion

What it does

Unlimited OCR Works is a Baidu-built document parsing model that ingests single images or multi-page PDFs and emits structured text in a single inference pass. It runs via Hugging Face Transformers or an SGLang server and offers two single-image modes—gundam and base—alongside a multi-page pipeline that rasterizes PDFs into images for processing. The README focuses almost entirely on inference setup and leaves the underlying architecture unexplained.

The interesting bit

The model fights long-horizon hallucination and repetition with a custom DeepseekOCRNoRepeatNGramLogitProcessor that bans 35-grams inside a sliding window—128 tokens for single images, 1024 for multi-page. That is the actual mechanism behind the “unlimited” claim; it is less about infinite context and more about not getting bored halfway through a book.

Key highlights

• Supports single-image gundam mode (cropped, 640 px) and full base mode (1024 px).
• Multi-page and PDF parsing via infer_multi, though PDFs must first be converted to images with PyMuPDF.
• Ships with a custom no-repeat n-gram logit processor to prevent degenerate loops during long outputs.
• Serves an OpenAI-compatible API through SGLang for batch or streaming use.
• Explicitly positioned as a successor to DeepSeek-OCR, borrowing ideas from PaddleOCR.

Caveats

• The README claims to push past DeepSeek-OCR but offers no benchmarks or side-by-side comparisons to prove it.
• SGLang setup instructions list conflicting kernels versions (0.9.0 in prose, 0.11.7 in the code block).
• PDF support is indirect: pages are rasterized to PNGs at a chosen DPI rather than parsed natively.

Verdict

Worth a look if you are building pipelines that need to OCR hundred-page PDFs in a single pass and can tolerate rasterized inputs. Skip it if you need native PDF extraction, quantified accuracy claims, or a detailed model card.