Indra OCR: Tiny 1M-parameter model reads PDFs and scans on CPU alone

What this is about

Datalab, a document processing company, has released the model Indra OCR on Hugging Face. OCR stands for optical character recognition: software that reads text out of images or scanned PDFs. Indra OCR is extremely small. According to the Model Card, the main model has around 1 million parameters (parameters are, roughly speaking, the dials a model tunes during training). For comparison, today's language models often have billions. The model runs on ordinary processors, with no GPU at all.

It converts PDFs and images to Markdown, plain text, or JSON with bounding boxes. According to the maker, it reads handwriting, more than 80 languages, and pages that are rotated or scanned at a skew. You install it with pip install indra-ocr.

What the model can do

• Layout: For every line and block, it returns a box, a label and a reading order. It detects headings, lists, tables, captions, footnotes and page headers.
• Real scans: Pages that are sideways or upside down, tilted by up to 45 degrees, or have light text on a dark background.
• Digital PDFs: If a PDF has a text layer, Indra OCR reads it in any script and applies OCR only to what's missing, such as text inside images.
• Output: Markdown with headings and lists. Tables come out as HTML. The JSON contains lines, boxes, layout blocks, confidence scores and timings.
• Math: Formulas come out as plain text by default. With --math, a small classifier detects formula lines and a second recognizer writes them as LaTeX.

Technology and languages

Under the hood, four recognizers work together, and a line classifier sends each line to the right one:

• rec.onnx: Latin script (about 50 languages), Cyrillic, Greek and handwriting in Latin script
• rec_cjk.onnx: Chinese, Japanese, Korean
• rec_rtl.onnx: Arabic, Persian, Urdu, Pashto, Kurdish, Hebrew
• rec_world.onnx: Devanagari, Bengali, Tamil, Thai, Khmer, Georgian, Armenian, Ethiopic and more

If the classifier is unsure, the main recognizer reads the line too, and the more confident result wins. The default models are about 76 MB in total: 30 MB of networks and 46 MB of language models for uncertain lines. The runtime is ONNX Runtime (a format and engine for running models efficiently), plus OpenCV and pypdfium2, but no PyTorch.

In tests on printed pages, the maker says 83 of 95 languages reached a character F1 of at least 0.90. Vertically set Chinese and Japanese pages are supported as well.

Numbers and benchmarks

The figures come from the maker. OmniParseBench (16,288 tests across 2,937 pages) also comes from Datalab, and the Indra adapter is part of the benchmark.

PP-StructureV3 reads text better than Indra without --math (55.6 vs. 46.7), but takes about 190 times as long per page.

For throughput on an AMD EPYC Genoa server (no GPU), Indra OCR reaches 6.2 pages per second on 4 cores and 43.9 on 32 cores. Tesseract manages 2.92 and PaddleOCR's text models 4.28. On other server types, Datalab measured 25 to 51 pages per second at 32 cores. On olmOCR-Bench, it beats Tesseract 4 on headers and footers (91.2 vs. 35.5) and tables (60.5 vs. 1.0), among other categories, and ties on multi-column text.

Limitations

Datalab is open about the weak spots. The model is built for fast processing of clean print:

• Handwriting: just 25.2% on the handwriting tests. Roughly a third of words are misread, with cursive and faint pencil doing worst.
• Complex tables: merged cells or tables without ruling lines can produce wrong rows and columns.
• Old or poor scans: only 29.5% on the degraded-scan tests.
• Diacritics: In Czech, Slovak, Turkish, Lithuanian and others, a similar-looking character can replace the correct one. According to the maker, English, French, German and Spanish are not affected.
• Very wide banner pages, technical drawings and newspapers with nested article order don't work well. Pages over 100 megapixels at 200 dpi are rendered at a lower resolution.

License and use

The code is under Apache 2.0. The weights use a modified OpenRAIL-M license. They are free for research, personal use and startups with under $20 million in funding or revenue, and they may not be used to compete with the Datalab API. For other cases, Datalab sells licenses. For top accuracy, the company points to its larger model Chandra on its cloud platform.

What you need to try it

You don't need a GPU. An ordinary computer with Python 3.10 or newer is enough (tested on Linux x86_64 and aarch64, and on macOS with Apple Silicon). The models are about 76 MB and are downloaded on first launch. More cores mean more pages per second. On 4 cores, the maker quotes around 210 ms per page. Run indra-ocr-viewer to start a local web viewer. Your files stay on your machine.

Our take

Indra OCR shows that many everyday documents don't need a large vision-language model. If you want to convert lots of clean PDFs or scans to Markdown, for example for search or RAG pipelines (systems that make documents searchable for AI answers), you get a very fast tool without a GPU. The caveats: the benchmarks come from the maker itself, and the model is weak on handwriting, complex tables and poor scans. The license isn't fully open, and the company is promoting its paid Chandra at the same time. Test it on your own files to see how well it handles German documents in practice.