Indra OCR: 1M-parameter open model hits 49 pages/sec on CPU, 11-17x faster than Tesseract
VikParuchuri · x · 2026-10-17
Datalab (VikParuchuri's team, of Marker fame) released Indra OCR, an open-source 1M-parameter OCR model that processes 49 pages per second on CPU, turning PDFs and images into markdown, plain text, or JSON with boxes.
Key numbers:
- 49 pages/sec throughput, 11-17x faster than Tesseract and PaddleOCR; 230ms per page latency on 4 x86 vCPUs (12-24x faster)
- Accuracy: 41.3 on OmniParseBench text vs PaddleOCR PP-OCR (35.8) and Tesseract (19.3); headline 62.1 beats PaddleOCR's full PP-StructureV3 pipeline (58.4)
- Bounding boxes per line and layout block with labels and reading order; digital PDFs read the native text layer and OCR only what's missing
- 18MB of ONNX models, no PyTorch needed; pip install indra-ocr
Hard cases (handwriting, math, complex tables, degraded scans) remain for their paid platform's Chandra model.
Related event: Indra: A 1M-Parameter Open OCR Model 15x Faster Than Tesseract(3 posts)→
More from Multimodal
- Turn your kid's stuffed animal into an AI cartoon with Nano Banana + Dreamina — AIandDesign · 2026-10-20
- MiniMax H3 tops AA video leaderboard as FLUX 3 Max debuts at No.2 with half the price — ArtificialAnlys · 2026-10-20
- DDSP Study: Magnitude Compression in Spectral Losses Helps Only If It Matches the Informative Region — kastnerkyle · 2026-10-20
- Seedance 2.5 object replacement applied to a Fast & Furious clip — Ror_Fly · 2026-10-20
- Same video model, 3x price gap: users call out Higgsfield's pricing — videorouter · 2026-10-20
- GPT-6 Astra in spatial XR: Jarvis-style live Earth demo on a PICO device — Scobleizer · 2026-10-20