Dots OCR

dots.ocr is a vision-language model (VLM) for multilingual document layout parsing that unifies layout detection and content recognition within a single 1.7B-parameter LLM foundation, outputting structured JSON with elements sorted in human reading order. It achieves state-of-the-art performance on the OmniDocBench benchmark, with the lowest overall Edit distance scores (0.125 EN, 0.160 ZH), the highest table TEDS scores (88.6 EN, 89.0 ZH), and the best reading order Edit scores (0.040 EN, 0.067 ZH) among all compared methods, including general VLMs like Gemini2.5-Pro and Doubao-1.5. Its formula recognition results (0.329 EN, 0.416 ZH Edit) are competitive with much larger models. On the new XDocParse benchmark spanning 126 languages, dots.ocr achieves state-of-the-art performance with approximately 10% relative improvement over prior methods. The model also demonstrates robust parsing for low-resource languages, achieving decisive advantages on an in-house multilingual benchmark. Its unified architecture allows task switching via prompt changes, providing faster inference than models built on larger foundations. Additional evaluation on olmOCR-Bench yields an overall score of 79.1, with 82.1 on Arxiv Math, 64.2 on Old Scans Math, and 88.3 on Table Tests.

Benchmark Results

End-to-end evaluation on OmniDocBench (lower Edit is better, higher TEDS is better):

Method	Overall (EN)	Overall (ZH)	Text (EN)	Text (ZH)	Formula (EN)	Formula (ZH)	Table (EN)	Table (ZH)	Read Order (EN)	Read Order (ZH)
dots.ocr	0.125	0.160	0.032	0.066	0.329	0.416	88.6	89.0	0.040	0.067