Datalab Drops a 1M-Parameter OCR Model, Claims 49 Pages a Second on CPU
Indra OCR is available as a local package for PDFs and images; its speed and accuracy comparisons are Datalab's own benchmark results.
A small recognizer for bulk documents
Datalab released Indra OCR on October 13, pitching a downloadable way to turn scanned pages and digital PDFs into Markdown, plain text or structured JSON on CPUs. Its main text recognizer has roughly one million parameters; the complete package also handles page orientation, line boxes, layout labels and reading order. Datalab says the model files total about 18 MB and provides a Python package and command-line tool. The release was discussed publicly on Hacker News the same day.
Fast in Datalab’s tests, with clear limits
The headline 49.4 pages per second comes from Datalab’s own run on 32 CPU cores, processing the same 300-page benchmark set with one worker per core. It is not a per-core rate or an independent result. Datalab also reports a shorter single-page latency and compares its text and layout scores with Tesseract and PaddleOCR under specified settings. The company says handwriting, degraded scans and complex tables remain weak spots; some scripts work only through a digital PDF’s existing text layer. The repository lists Apache 2.0 for code, while the model card lists an OpenRAIL license for the weights, so the two licenses should not be treated as identical.