olmOCR

Use it when

• You have scanned or image-based PDFs with multi-column layouts, equations, tables, or handwriting and need Markdown or text in natural reading order.
• You need to process a large document corpus on your own GPU or compatible hosted inference, with optional S3-coordinated workers, instead of using a conventional OCR API.

What it solves

Not the fit when

• CPU-only local inference: the documented local path requires a recent NVIDIA GPU with at least 12 GB VRAM; remote --server mode shifts the compute and cost elsewhere.
• A fixed per-page price assumption: the README estimates under 200 USD per million pages for its tested setup, but actual cost varies by hardware and hosting.
• Born-digital PDFs where a lightweight text extractor already suffices.
• Audio or video transcription and general web scraping.
• Unreviewed extraction in high-stakes workflows where OCR errors can change legal, medical, financial, or safety meaning.
• media-transcript-quality
• embedding-recall-poor
• messy-json-yaml

Install

pip install olmocr[gpu] --extra-index-url https://download.pytorch.org/whl/cu128 (or pip install olmocr for remote-server use; Docker image alleninstituteforai/olmocr:latest-with-model also available)

Invoke

olmocr ./localworkspace --markdown --pdfs document.pdf; add --server http://host:8000/v1 --model allenai/olmOCR-2-7B-1025-FP8 to use a remote vLLM or OpenAI-compatible endpoint.

Alternatives

No reviewed alternatives recorded yet.