Indra OCR parses documents with a 1-million-parameter recognizer

Datalab has released Indra OCR, a compact document-conversion stack that runs on CPUs and turns PDFs or images into Markdown, plain text, or structured JSON. Its output can include line-level bounding boxes, labeled layout blocks, reading order, and confidence scores.

Modern document parsers often use vision-language models that render each page as an image and generate text token by token, increasing memory use and latency. Indra targets high-volume ingestion jobs such as preparing documents for retrieval-augmented generation, indexing scanned archives, and routing difficult pages to larger models.

The repository code uses the Apache 2.0 license. The model weights use a modified OpenRAIL-M license that permits free research use and free use by startups with less than $20 million in funding or revenue. Organizations outside those terms should review the model license before deployment.

49 pages per second, under test

On a 32-core AMD EPYC system, Datalab reports processing 300 OmniParseBench pages at 49 pages per second. A 4-vCPU configuration averaged roughly 230 milliseconds of latency per page.