dots.ocr-1.5 turns charts and diagrams into clean SVG code

rednote-hilab just dropped dots.ocr-1.5 on Hugging Face, a compact 3B-parameter multimodal OCR model built for document parsing.

What it does

The model handles multilingual document extraction and claims SOTA performance across virtually any human script.

Its standout feature is converting structured graphics into code. As noted by xeophon on X, dots.ocr-1.5 excels at turning charts and diagrams directly into clean, editable SVG code. Instead of spitting out fragmented text snippets or flattening diagrams into static bitmaps, it preserves the underlying layout and structure as vector graphics.

The model weights are available to download on Hugging Face, and the 3B footprint makes it relatively accessible to run.

Why it matters

If your pipeline regularly parses technical papers, financial reports, or PDFs packed with visual figures, this is well worth testing.

We think the direct-to-SVG trick is especially practical. Rather than relying on a massive general-purpose vision model, using a targeted 3B model to reconstruct vector markup directly cuts out a huge chunk of downstream post-processing for document parsing pipelines.