Unlimited OCR: One-shot long-horizon parsing
Points and comments are a snapshot, not live.
Baidu's Unlimited-OCR uses Reference Sliding Window Attention to parse long documents in one shot.
Unlimited-OCR extends Deepseek-OCR with a memory-efficient transformer architecture. Reference Sliding Window Attention splits focus into a global reference (full original image) and a local generation window of ~128 words, limiting KV-cache growth. It is released on GitHub and Hugging Face Spaces. Inference supports single images in two modes (gundam: 640px with cropping; base: 1024px without) and multi-page/PDF parsing at 1024px. A paper is available on arXiv.
What commenters are saying
Strong disagreement with the claim that OCR is solved. Commenters report that traditional OCR fails on complex layouts, non-Latin scripts (CJK, Arabic, Thai), handwriting, and mixed-content documents. Vision-LLMs improve layout understanding and script variance but introduce harder-to-detect failure modes and higher costs. Several users name specific tools: marker (with --force-ocr) for complex documents, Azure Document Intelligence and AWS Textract for commercial options, and PaddleOCR for open-source CJK recognition.