Datalab · depuis 13. Oktober 2026 · 3× · vu le 18 oct. 2026
48
Momentum
Indra OCR is an open-source OCR package from Datalab whose core recognition model has about 1M parameters. It converts PDFs and images into Markdown, plain text or JSON with bounding boxes and layout blocks, and runs on CPU without GPU or PyTorch via ONNX Runtime. Datalab reports 49 pages per second on 32 CPU cores; independent tests on other hardware measure notably lower figures.
Historique du momentum
21.07.19.10.
Fonctionnalités
| Key Benchmark (%) | OmniParseBench overall 62.1 (text 41.3; tables 56.4; layout 88.6); text vs. Tesseract 19.3 and PaddleOCR PP-OCR 35.8 (per Datalab) |
| License | Code: Apache 2.0; model weights: modified OpenRAIL-M (free for research, personal use and startups under $20M funding/revenue; not for uses competing with Datalab's API) |
| Multimodality | Not multimodal in the LLM sense: pure OCR/document processing; output text, Markdown, JSON with layout; no handwriting, no vertical text |
| Platform | CPU-optimized (ONNX Runtime, no PyTorch, no GPU needed); Python 3.10+, pip install indra-ocr |
| Release Date | October 13, 2026 |