What is Indra OCR?
Indra OCR is an open-source OCR model and pipeline from Datalab, the company behind Marker and Chandra. It turns PDFs and images into Markdown, text or JSON with line and block boxes, labels and reading order, and runs on ONNX Runtime with no PyTorch and no GPU. The main Latin recogniser has about 1.25M parameters; all default models together are 7.3M parameters and a 76 MB download.
It reads the text layer of digital PDFs and OCRs only what is missing, handles rotated and skewed scans, covers more than 80 languages through four script-specific recognisers, and has an optional math mode that writes LaTeX. It installs with pip install indra-ocr and includes a local web viewer.
The code is Apache-2.0, but the weights use a modified OpenRAIL-M licence that is free only for research, personal use and organisations under $20M in funding or revenue, and bars use that competes with Datalab's API. The repo, created on 13 October 2026, has about 1,530 GitHub stars; the latest release is v1.1.2 (20 October 2026).
Is Indra OCR worth it?
Indra's point is throughput per pound. On Datalab's figures it reads about 44 pages a second on 32 CPU cores and takes a median 212 ms per page on four, against 5,496 ms for Tesseract 5 and about 41 seconds for PaddleOCR's full PP-StructureV3 pipeline on the same pages. For a team OCRing millions of pages on ordinary cloud instances, that turns a GPU budget line into a CPU one. Every number here is the maker's own, but the harness and reproduction commands are in the repo.
Quality is good rather than leading. On Datalab's OmniParseBench it scores 66.7 overall (71.4 with math mode) against 58.4 for PP-StructureV3, and its layout score of 92.2 is the strongest part. On plain text, though, PP-StructureV3 still wins without math mode (55.6 against 46.7). The README is unusually frank about the rest: complex tables with merged cells, handwriting (25.2% on the handwriting tests), degraded scans (29.5%) and some heavily accented Latin languages such as Czech and Turkish read poorly. It points those documents to Datalab's paid Chandra model, which is the commercial logic of the release.
The licence is the decision that matters for a business. The code is Apache-2.0, but the weights are under a modified OpenRAIL-M that is free for research, personal use and startups under $20M in funding or revenue, and forbids use that competes with Datalab's API. A larger company, or anyone building an OCR service, needs a commercial licence. Datalab licenses Marker's weights in a similar way, and the terms rule Indra out as a drop-in Tesseract replacement for many enterprises.
The benchmark it leads on was also built by Datalab, and the repo is less than three weeks old. Against Marker, Indra trades accuracy on hard layouts for speed and a far smaller footprint; against Tesseract it is faster and much better at layout and tables, with a licence Tesseract does not impose.
Where it earns its place
A data engineer at a 30-person legal-tech startup
OCR a 2-million-page archive of scanned contracts for search
Run indra-ocr over the archive on a 32-core instance; at Datalab's 44 pages a second that is roughly 13 hours of compute. The weights are free while the company stays under $20M in funding or revenue.
An ML engineer building a RAG pipeline
Get chunkable Markdown with layout labels from mixed PDFs
Use convert_many with one worker per core; page.blocks returns headings, lists, tables and captions in reading order, and digital PDFs skip OCR via the text layer. Route handwritten or degraded pages to a heavier model.
A procurement lead at a large enterprise
Decide whether Indra can replace Tesseract in production
Check revenue and funding against the $20M ceiling first; above it, or for any OCR-as-a-service product, the weights need a commercial licence from Datalab's pricing page.
Before you commit
- Setup
- pip install indra-ocr on Python 3.10+, then indra-ocr file.pdf; the first run downloads about 76 MB of models from Hugging Face. Tested on Linux x86_64 and aarch64 and Apple-silicon macOS.
- Integrations
- CLI, Python API (load, convert, convert_many, ocr on numpy arrays), Markdown, text or JSONL output, a local web viewer on port 8765, and an optional Tesseract comparison.
- Data and privacy
- Runs locally; files stay on your machine and only the model download touches the network. Code Apache-2.0; weights under a modified OpenRAIL-M with a $20M funding or revenue limit and a no-competing-API clause. No security advisories published.
- Cost at scale
- CPU only: Datalab measures 6.2 pages a second on 4 cores and 43.9 on 32. Larger companies need a commercial weights licence; Datalab's hosted API offers about 5,000 free pages a month and SOC 2 Type 2 for harder documents.
- Limits
- Complex tables, handwriting, degraded scans and scripts outside the list read poorly; math is plain text unless --math is on, which raises median latency to 357 ms; about 18 open issues.
What’s good
- About 212 ms a page on four CPU cores; no GPU or PyTorch
- Layout blocks, reading order and boxes in JSON as well as Markdown
- More than 80 languages; reads digital PDF text layers first
- Pip install, CLI, Python API and local viewer
What’s not
- Weights licence caps free commercial use at $20M funding or revenue
- Weak on handwriting, degraded scans and complex tables
- Benchmarks are Datalab's own, on Datalab's benchmark
- Repo is under three weeks old
Best for
Startups and research teams that need to OCR large volumes of clean printed documents cheaply on CPUs.
Indra OCR alternatives
Indra OCR: common questions
- Is Indra OCR worth it?
- Fast enough on plain CPUs to change the economics of bulk OCR, provided your documents are clean print and your company fits under the licence's $20M line. We rate it 4.0 out of 5.
- How much does Indra OCR cost?
- Code Apache-2.0; weights free under a modified OpenRAIL-M for research, personal use and companies under $20M funding or revenue; commercial licence from Datalab otherwise.
- What are the downsides of Indra OCR?
- Weights licence caps free commercial use at $20M funding or revenue; weak on handwriting, degraded scans and complex tables; benchmarks are Datalab's own, on Datalab's benchmark; repo is under three weeks old.
- Who is Indra OCR best for?
- Startups and research teams that need to OCR large volumes of clean printed documents cheaply on CPUs.
- How much does Indra OCR cost for a team or in production?
- CPU only: Datalab measures 6.2 pages a second on 4 cores and 43.9 on 32. Larger companies need a commercial weights licence; Datalab's hosted API offers about 5,000 free pages a month and SOC 2 Type 2 for harder documents.
- Is Indra OCR safe for company data?
- Runs locally; files stay on your machine and only the model download touches the network. Code Apache-2.0; weights under a modified OpenRAIL-M with a $20M funding or revenue limit and a no-competing-API clause. No security advisories published.
- What are the best Indra OCR alternatives?
- The closest alternatives we have reviewed are Marker, Docling, Ocrs and deepseek-ocr.rs.
Answers checked Oct 31, 2026.
Make Indra OCR? Show this rating on your site
