Datalab's Indra: open OCR that runs on a CPU at 44 pages/sec
A 76 MB ONNX model turns PDFs into markdown with layout boxes — ~15x Tesseract across 95 languages, but the open weights carry a non-compete license.
Copy markdown
1M params, zero GPUs
Indra is a ~1M-parameter recognizer — about 76 MB of ONNX weights — that runs on CPU with no PyTorch and no CUDA. Datalab clocks ~44 pages/sec on 32 cores and ~212 ms/page on four, roughly 15x Tesseract on the same machine.
Structured output, not just text
`pip install indra-ocr` then `indra-ocr paper.pdf` emits markdown, text, or JSONL with per-line bounding boxes, layout labels (headings, tables, lists) and reading order — a drop-in front end for a RAG ingestion pipeline. `indra-ocr-viewer` renders the boxes locally.
The benchmark, and where it breaks
It scores 66.7 on OmniParseBench (71.4 with `--math`) versus 58.4 for PaddleOCR's PP-StructureV3, and reads 83 of 95 tested languages at F1 ≥ 0.90. It's weak on complex tables, handwriting and degraded scans, so vet those before production.
Read the weights license first
The code is Apache 2.0, but the weights ship under a modified OpenRAIL-M: free for research, personal use and startups under $20M in funding or revenue — and explicitly not for building a product that competes with Datalab's own OCR API.