dots.ocr-1.5

Overview

dots.ocr-1.5 is a 3B-parameter OCR model from rednote-hilab, made of a 1.2B vision encoder and a 1.7B language model. It splits a page into labeled blocks and returns the text of each one in reading order. The team aimed it at multilingual documents, and the examples on its model card include pages in Tibetan, Kannada, Russian, Dutch and Traditional Chinese.

rednote-hilab reports an olmOCR-Bench overall score of 83.9, up from 79.1 for the first dots.ocr, and a text edit distance of 0.031 on OmniDocBench v1.5, where lower is better.

The checkpoint here is kristaller486's re-upload of the original MIT release from ModelScope. rednote-hilab has since renamed the model dots.mocr.

What you can build

• Resume intake. On the sample two-column resume, the main column is read first and the sidebar after it.
• Searchable text for scanned reports and papers, with tables kept as HTML and formulas as LaTeX.
• Intake for archives where one batch holds pages in several scripts.
• Segmentation checks: the annotated image shows where a page was split wrongly before you trust the text.
• Text and layout from web screenshots, or from photos of signs and packaging.

Inputs

Upload one file per run: a PDF, or an image in JPG, PNG, WebP, BMP, HEIC, TIFF or AVIF format. A PDF is processed page by page.

Prompt Mode picks the task:

• prompt_layout_all_en (default): every block with its box, category and text.
• prompt_layout_only_en: boxes and categories, no text.
• prompt_ocr: the text alone, without page headers and footers.
• prompt_web_parsing: the layout of a webpage screenshot.
• prompt_scene_spotting: locates and reads text in ordinary photos.

The Prompt field comes pre-filled with the instruction behind the default mode. Whatever it holds replaces the instruction sent to the model, while Prompt Mode still decides how the reply is processed and which files come back. You can replace it with your own instruction.

Min Pixels and Max Pixels bound the resolution each image is resized into before reading. Both accept 3,136 to 11,289,600 pixels, and the defaults sit at the two ends of that range.

Outputs

In the default mode each image, or each PDF page, gives an annotated image and a text file. In prompt_layout_only_en you get only the annotated image, and the other three modes return only a text file with the model's raw reply. The image is the original page with numbered, color-coded boxes in reading order, each labeled with one of 11 categories such as Title, Section-header, Table, Formula or Picture.

In the default mode the text file is Markdown in reading order. Titles and section headers become headings, tables are written as HTML and formulas as LaTeX. Picture regions are not read. Each one is cropped and embedded in the file as an inline base64 PNG image. Nothing is translated.

Recommended settings

Use the default mode for full documents and prompt_ocr when you only need the words. Empty the Prompt field before switching modes. Keep Max Pixels at its ceiling for dense pages and small print. A custom prompt replaces the preset instruction entirely. The two layout modes still expect the JSON layout back to draw boxes, so run free-form prompts under prompt_ocr, where the reply is saved unchanged.

Limitations

• rednote-hilab says complex tables and mathematical formulas remain difficult for a model this compact.
• Parsing failures are rarer than in dots.ocr but still happen.
• Chart-to-SVG conversion, which the card describes, is not one of the preset modes here. The team says its SVG quality is not yet where it wants and released a separate dots.ocr-1.5-svg model for it.

Safety & compliance

The weights are released under the MIT license. A dots.ocr license agreement ships alongside them and adds acceptable-use terms, including no extraction of sensitive personal data without consent or a lawful basis, and no digitizing of publications or other copyrighted material without permission.