indra-ocr

indra-ocr

A tiny OCR model that runs on a CPU and turns PDFs and images into markdown, text or JSON with boxes. Reads handwriting and more than 80 languages. For developers and data teams.

Made by datalab-to on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: indra-ocr is a compact OCR model from Datalab, about 1.3 million parameters by Hugging Face's count. It runs on a plain CPU through ONNX Runtime, with no PyTorch and no GPU.

It turns PDFs and images into markdown, plain text or JSON with boxes. The card says it reads handwriting, more than 80 languages and pages that are sideways, upside down or tilted.

Who it suits: Developers and data teams who need text out of scans and documents on ordinary hardware, and anyone who wants layout boxes along with the words.

Industry | Who uses it | How they use it | Example
Document processing | Data engineers | Convert batches of PDFs to markdown | Turn a folder of reports into text files
Software | Developers building apps | Call a command line tool inside a pipeline | Get JSON with boxes and layout labels from a scan
Records and archives | Archivists and office teams | Read scanned letters and forms | Pull text from a handwritten page

Skip it if you need the best accuracy on handwriting, math or messy tables. The card sends those jobs to the makers' managed platform.

Where to use it:

• A command line, after installing the Python package
• Python programs and document pipelines
• A local viewer that ships with the package
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Scanned documents | Office teams | Markdown from each page
Layout work | Developers | Position boxes, type labels and a reading sequence for lines and blocks
Digital PDFs | Data teams | The text layer read in any script, with OCR only where text is missing

How to set it up:

• Python and pip
• An ordinary CPU, since no GPU is needed
• About 76 MB of models, which download on the first run

1. Install the package.
2. Run the command on a PDF or image.
3. Add options to choose the output format or an output folder.

pip install indra-ocr

indra-ocr paper.pdf
indra-ocr *.pdf --out results/
indra-ocr scan.png --format json
indra-ocr-viewer

How to use it: Give the command a file and it prints markdown. Use the out option to write one markdown file per input, and the format option to get JSON with lines, boxes, layout blocks and confidences.

Command | What it does
indra-ocr paper.pdf | Prints markdown for the document
indra-ocr .pdf --out results/ | Writes one markdown file per input, using all cores
indra-ocr scan.png --format json | Returns lines, boxes, layout blocks and confidences
indra-ocr-viewer | Opens a viewer on your own machine

Example: You run the tool on a scanned letter. It finds each line, reads each one with the recognizer for its script, and gives you markdown, or JSON if you asked for that.

For newcomers and experts: If you are new to OCR, run the tool on one clear page first and read the markdown. Then try a tilted or handwritten page to see where it struggles.

If you know the field, the makers report speed numbers on CPU and a lower handwriting score than their managed model. Their handwriting test result is 25.2 percent, so check that against your own pages.

Install the package
Run it on a PDF or image
Lines and layout blocks are found
Each line is read in its script
Markdown, text or JSON comes out

Good to know:

• The scores below are the makers' own, measured on their benchmark
• The makers report roughly 44 pages each second when 32 CPU cores are used, and roughly 210 ms per page with 4 cores
• Hugging Face lists 7 quantized and 2 fine-tuned versions made by other people

Title: Text score on the makers' benchmark
Indra OCR: 46.7 points
PaddleOCR PP-OCR: 35.8 points
Tesseract: 19.3 points

• License: the code is Apache 2.0, and the weights use a modified OpenRAIL-M license, so read it before commercial use
• Cost to run: none named for the local tool
• Data it reaches: the files you download and the documents you scan
• Not for: top accuracy on handwriting, math and degraded scans

On Hugging Face, 13 November 2026

• Task: image to text
• Library: onnx
• Parameters: 1.3M
• Weights: about 5 MB in F32; about 1 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 8,624
• Likes: 529
• License: openrail
• Languages: en, fr, de, es, it, pt, nl, pl
• Quantized versions: 7
• Fine-tunes: 2
• Tags: onnx, safetensors, indra-ocr, markdown, layout, document-parsing, handwriting, multilingual, image-to-text

Links

• Model: https://huggingface.co/datalab-to/indra-ocr
• Files: https://huggingface.co/datalab-to/indra-ocr/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download datalab-to/indra-ocr

Then follow the usage in its model card. Nothing in this kit runs by itself.