Indra OCR is Datalab's CPU-first OCR engine, built around a recognizer (rec.onnx) of about 1M parameters that the README says reads up to 49 pages per second on CPU. It turns PDFs and images into markdown, plain text, or JSON containing a bounding box for every line and every layout block, with block labels such as headings, lists, tables, captions and footnotes, plus reading order. For digital PDFs it reads the file's own text layer in any script and OCRs only what is missing; for scans it covers Latin (about 50 languages), Cyrillic, Greek, Devanagari, Bengali and Chinese. The package installs with pip install indra-ocr, runs on ONNX Runtime with no PyTorch, and downloads about 18 MB of models from Hugging Face on first use. It ships a CLI (indra-ocr paper.pdf, with --format md|text|json, --mode auto|ocr|text, --math for LaTeX), a Python API (indra_ocr.load and convert_many) and a local web viewer. The README reports benchmarks on OmniParseBench's 2,937 pages on x86 with 4 vCPUs: a headline score of 62.1 and a text score of 41.3 in default auto mode, against 35.8 for PaddleOCR's PP-OCR text models and 19.3 for Tesseract 5, at a median of 228 ms per page; with --math the headline rises to 66.8 and text to 57.5 at roughly 360 ms. Throughput on 32 cores is given as 49.4 pages per second. The authors list limits themselves: complex tables with merged cells, handwriting, degraded scans, vertical text, and scripts outside the OCR list (Arabic, Hebrew, Thai, Japanese kana, Korean Hangul) work only through a PDF's text layer. The code is Apache 2.0; the model weights use a modified OpenRAIL-M license that the README describes as free for research, personal use and startups under $20M funding or revenue, and not usable competitively with Datalab's API.