FULL STORY

Indra: The 1M-Parameter OCR Model's Rise and Update

Datalab open-sourced Indra, a 1M-parameter OCR model running dozens of pages per second on CPU, followed by wide coverage and a 1.1 update.

2026-10-14 ~ 2026-10-20 · 3 episodes · 13 posts

Episode 1 · Datalab Open-Sources Indra, a 1M-Parameter OCR Model Hitting 49 Pages/sec on CPU (2026-10-14, 6 posts)

Datalab founder Vik Paruchuri open-sourced Indra, a tiny OCR model with only 1 million parameters that runs at 49 pages per second on CPU, claimed to be 11x faster than PaddleOCR and 17x faster than Tesseract. Weights are publicly available on Hugging Face with one-step pip installation, potentially lowering the compute barrier for large-scale document processing.

Confirmed

  • Indra has 1M parameters and reaches 49 pages/sec on CPU; the author's benchmarks show it 11x faster than PaddleOCR and 17x faster than Tesseract, backed by OmniParseBench
  • Supports 60 languages across 6 scripts, plus a math mode; targeted at low-latency CPU scenarios such as RAG ingestion, search indexing, and LLM front-end triage
  • Available via Hugging Face weights (datalab-to/indra-ocr, openrail license, 18MB ONNX model), plus code and a blog post
  • Easy to use: pip install indra-ocr; a local test tool indra-ocr-viewer lets users check recognition speed and parameter settings
  • The author says Indra will keep improving, with plans for more languages and better accuracy

Why it matters

  • Traditional OCR is slow on CPU; Indra delivers an order-of-magnitude speedup with a minimal parameter count, lowering compute costs for document digitization, RAG ingestion, and search indexing at scale
  • Its 18MB size and pip-installable deployment make it easy to embed in production pipelines

Episode 2 · Indra: A 1M-Parameter Open OCR Model 15x Faster Than Tesseract (2026-10-16, 3 posts)

Datalab released Indra, an open-source OCR model with only 1M parameters that runs 15x faster than Tesseract on CPU while being more accurate, processing 49 pages per second and trending on Hugging Face.

Episode 3 · Datalab Releases Indra OCR 1.1: 1M-Parameter Model Runs 44 Pages/sec on CPU (2026-10-20, 4 posts)

Datalab released Indra OCR 1.1, a 1M-parameter OCR model that processes 44 pages/sec on CPU with no GPU required; it scores 71.4 on OmniParseBench with the --math flag, beating Tencent's 1.1B HunyuanOCR and approaching DeepSeek OCR 2.