Datalab released Indra, a 1M-parameter OCR model that beats Tesseract

Pulling text out of a scanned PDF is the step that quietly costs a team more than it should, because the Tesseract baseline is slow and the modern 1B-parameter models want a GPU. Datalab's new OCR model, Indra, sits in the gap. The README on the datalab-to/indra GitHub repository says Indra has about 1.25 million parameters, runs through ONNX on a plain CPU, and reports 15 times the throughput of Tesseract on the same hardware. The repository was created on 13 October 2026 and has 890 GitHub stars on 22 October.

What Datalab reports

Datalab's own benchmark on OmniParseBench puts Indra at 66.7 overall and 71.4 with the --math flag turned on. The text sub-score is 46.7, against Tesseract at 19.3 and PaddleOCR at 35.8 on the same test, and 62.7 under --math. Layout detection scores 92.2 and tables 61.1. The company says the small recognizer runs at 43.9 pages per second on 32 cores and about 210 milliseconds per page on 4 cores, which fits a small virtual machine rather than a GPU host.

The model file is listed at about 76 MB, and the README names more than 80 languages, grouped into Latin script, Cyrillic, CJK, Arabic, Indic scripts and the rest, plus right-to-left and vertical layouts for the ones that use them. Code is Apache-2.0 licensed. The model weights sit under a modified OpenRAIL-M licence that Datalab says is free for research, personal use and startups under 20 million dollars of funding or revenue.

Where it is weaker, in the README's own words

The README is unusually direct about where Indra does not win. Handwriting scores 25.2 percent, which Datalab calls poor. Degraded or low-quality scans score 29.5 percent in its tests. The maths sub-score sits at 14.0 by default and reaches 55.8 only with the --math flag, so a document heavy in equations needs a separate pass. Complex tables with merged cells, scripts outside the listed set, and languages with heavy diacritics are also called out as cases where accuracy drops. A team pulling text out of printed invoices or digital-first PDFs will do very well; a team scanning old filings will not get all of it.

What this is doing to the OCR stack

An agent-based product that ingests documents often puts the OCR step on a model-sized line in its cost sheet, because the obvious upgrade from Tesseract was a billion-parameter vision model with its own GPU bill. A 1M-parameter model on an ONNX runtime is a different cost shape: a small container, the ability to run per-tenant without a GPU pool, and a throughput number that matches a batch pipeline rather than an interactive request. Datalab also runs a managed platform with a larger model called Chandra, so a team can start with Indra, measure what fails, and send the hard pages to the paid service without rewriting the pipeline.

Nothing in the Datalab README reports a comparison against the closed-source cloud OCR services, and the "close to 1B-plus parameter models in accuracy" line is a Datalab claim without a specific competitor number next to it. The right test is a few hundred of a team's own pages, judged against whatever is in use today.

Source

GitHub: datalab-to/indra. Datalab's benchmark figures are from its own README.