dots.ocr

dots.ocr is a multilingual document-parsing vision-language model from RedNote HiLab, released July 30, 2025. Built on a 1.7B language model, it unifies layout detection, text recognition, and relational understanding in one end-to-end VLM, avoiding brittle multi-stage OCR pipelines. It reports state-of-the-art results among comparable-size models on OmniDocBench, olmOCR-bench, and its own XDocParse benchmark spanning 126 languages.

The model also handles structured graphics (charts and diagrams converted to SVG), web screens, and scene text. In March 2026 the successor dots.ocr-1.5 was rebranded as dots.mocr. Weights are available on Hugging Face; check the model card for the current license before commercial use.