[Models](/models)/datalab-to/indra-ocr

Modified OpenRAIL-M (weights), Apache 2.0 (code)

# indra-ocr

by datalab-to

Datalab's CPU-only OCR model that turns PDFs and images into Markdown, text or JSON with boxes, in 80+ languages and handwriting.

500likes

About 76 MB of ONNX weightsparams

[HuggingFace](https://huggingface.co/datalab-to/indra-ocr)[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

Identifiers

Model ID

`datalab-to/indra-ocr`

## Deploy indra-ocr

Single-tenant

Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

[Read the customer-hosted guide](https://mixpeek.com/docs/customer-hosted/overview) [Talk to us about single-tenant](/demo)

## Overview

Indra OCR is an OCR model from Datalab, released on Hugging Face on October 10, 2026. It runs on ONNX Runtime with no PyTorch and no GPU, reads the text layer of digital PDFs directly and OCRs only what is missing, and returns a box, a label and a reading order for every line and block. Datalab reports 66.7 on its OmniParseBench, against 58.4 for PaddleOCR's PP-StructureV3 pipeline, and about 44 pages a second on 32 CPU cores. The weights use a modified OpenRAIL-M license with commercial limits, so read it before you ship.

## Architecture

A set of ONNX models, about 76 MB in total, run through ONNX Runtime on CPU. Each detected line goes to a recognizer for its script, and a layout model labels blocks (headings, lists, tables, captions, footnotes, page headers) and orders them for reading.

## Mixpeek SDK Integration

IngestRetrieve

\# Command line usage from the Indra OCR model card.
# pip install indra-ocr
#
# indra-ocr paper.pdf                  # markdown to stdout
# indra-ocr \*.pdf --out results/       # one .md per input, all cores
# indra-ocr scan.png --format json     # lines, boxes, layout blocks, confidences
#
# The first run downloads the models (about 76 MB) from Hugging Face.

## Capabilities

*   PDFs and images to Markdown, plain text or JSON with line and block boxes
*   80+ languages, including Cyrillic, Greek, CJK, Arabic, Hebrew and Indic scripts
*   Handwritten pages and forms
*   Pages rotated, upside down or tilted up to 45 degrees
*   Uses the PDF text layer when present and OCRs only the missing parts

## Use Cases on Mixpeek

Extracting text from large scanned archives on CPU machines before chunking for RAG

Getting line boxes and reading order to cite the exact region of a page in search results

Reading multilingual scans where a Latin-only OCR engine fails

Pre-filtering pages cheaply before sending hard ones to a larger vision-language model

## Benchmarks

Dataset

Metric

Score

Source

OmniParseBench

Overall score

66.7

https://huggingface.co/datalab-to/indra-ocr

OmniParseBench

Text score

46.7

https://huggingface.co/datalab-to/indra-ocr

OmniParseBench handwriting

Score

25.2%

https://huggingface.co/datalab-to/indra-ocr

## Performance

Input SizePDF pages and document images

GPU LatencyAbout 210 ms per page on 4 CPU cores; about 44 pages/s on 32 cores (Datalab's numbers)

GPU ThroughputBatch dependent

GPU MemoryModel dependent

We have not measured it. Scores and speeds are Datalab's own, from its model card.

## Common Pipeline Companions

[

xiaomi-research/Xiaomi-OCR-0

A 0.8B vision-language parser for pages that need charts and formulas

](/model/xiaomi-research/Xiaomi-OCR-0)[

PaddlePaddle/PaddleOCR-VL-1.5

A GPU document parser to compare against

](/model/PaddlePaddle/PaddleOCR-VL-1.5)[

athrael-soju/ColUltraVBERT

Retrieves the page image directly when you would rather skip OCR

](/model/athrael-soju/ColUltraVBERT)

## Frequently Asked Questions

### What is Indra OCR?

A CPU-only OCR model from Datalab that converts PDFs and images to Markdown, text or JSON with boxes. It runs on ONNX Runtime without a GPU and covers more than 80 languages and handwriting.

### How fast is Indra OCR?

Datalab reports about 44 pages a second on 32 CPU cores and about 210 ms per page on 4 cores, which it puts at 15 times Tesseract on the same machines. We have not measured it ourselves.

### Can I use Indra OCR commercially?

The code is Apache 2.0. The weights use a modified OpenRAIL-M license that the model card describes as free for research, personal use and startups under $20M in funding or revenue, and not usable in competition with Datalab's API. Larger companies need a license from Datalab.

## Explore on Mixpeek

[

Browse all models

Search the full model catalog

](/models)[

Feature Extractors

See how Mixpeek runs models as extractors

](/extractors)[

Mixpeek Vector Store

Store & search embeddings at scale

](/mvs)[

Pricing

Usage-based pricing for pipelines

](/pricing)[

All Curated Lists

Compare models, APIs & infrastructure

](/curated-lists)

### Specification

Organizationdatalab-to

Retriever\-

ParametersAbout 76 MB of ONNX weights

LicenseModified OpenRAIL-M (weights), Apache 2.0 (code)

Downloads/moN/A

Likes500

### Build a pipeline with indra-ocr

Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.

[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)