Baidu's Unlimited-OCR Scores 86 on Text and 0.97 on Formatting

Unlimited-OCR is a 3B MIT-licensed model for one-shot long-horizon document parsing, handling single images and multi-page PDFs with a 32,768-token context, and it has already passed 2.7 million downloads. Its ParseBench results deserve a careful read: 86.81 on text content, 46.17 overall, and 0.97 on text formatting. That is not a typo, and it draws the line sharply. For RAG ingestion where you only need the words, this is excellent and free. For invoices, tables or anything whose structure carries meaning, it is unusable. The download count suggests rather more people found the model than found the benchmark table.