dots.ocr

A single compact vision-language model replaces the usual OCR pipeline in dots.ocr, the document parser from Xiaohongshu's rednote-hilab. Instead of chaining a layout detector, a text recognizer, and a reading-order module, the model, built on a 1.7B LLM, treats all of it as one task: switching between layout extraction, plain text, or grounded output is just a prompt change. Results interleave bounding boxes with Markdown text, HTML tables, and LaTeX formulas across 11 layout categories, and the team's dots.ocr-bench stresses 1,493 PDF pages spanning about 100 languages, where the model leads open alternatives on low-resource scripts; it also posts top scores among compact models on OmniDocBench and olmOCR-bench. Weights are MIT licensed on Hugging Face, inference targets a CUDA GPU, and vLLM has carried official support since version 0.11, with Docker images and a plain transformers path for smaller setups. The same lab has since extended the line with dots.mocr, a 3B successor released with its own paper and weights in March 2026.