What it does

dots.ocr (now rebranded as dots.mocr) is a 3B-parameter vision-language model built for document understanding. It extracts text and structure from scanned pages, PDFs, web screenshots, and photographs — covering “virtually any human script,” including low-resource languages. It also converts charts, diagrams, and chemical formulas directly into SVG code rather than rasterizing them.

The interesting bit

The model punches above its weight class. On OmniDocBench and olmOCR-bench, dots.mocr outperforms specialized competitors like PaddleOCR-VL and HuanyuanOCR, and even edges closer to Gemini 3 Pro on some Elo-based rankings — all at 3B parameters versus Gemini’s undisclosed scale. The SVG generation trick is particularly unusual: it treats visual languages as code to reconstruct rather than pixels to describe.

Key highlights

• 3B parameters, available on HuggingFace with a smaller 1.7B “base” variant
• Benchmark-leading scores on olmOCR-Bench (83.9 overall) and OmniDocBench v1.5 (0.031 TextEdit distance)
• Handles multilingual text, web screenshots, scene text, and structured graphics in one model
• Outputs charts and diagrams as editable SVG code
• Live demo at dotsocr.xiaohongshu.com

Caveats

• The repo itself appears to be primarily a showcase page; model weights and inference code live on HuggingFace
• Some benchmark results (Gemini 3 Pro, PaddleOCR-VL-1.5) were obtained via API rather than reproducible local runs
• The README notes formula and table metrics were omitted from OmniDocBench reporting due to “high sensitivity to detection and matching protocols”

Verdict

Worth a look if you need offline, low-latency document parsing with broad language coverage — especially for mixed layouts with charts or non-Latin scripts. Skip it if you need the absolute highest accuracy regardless of cost; Gemini 3 Pro still leads on raw averages.