olmOCR

license

Apache-2.0(OSI)

source

public(allenai/olmocr)

core features withheld

no

olmOCR's code is Apache-2.0 and public, and the olmOCR-2 model weights it runs are published under Apache-2.0 as well. The whole pipeline installs and runs from the public repository, with no paid edition beside it.

• https://huggingface.co/api/models/allenai/olmOCR-2-7B-1025-FP8 recorded 2026-09-28

  Hugging Face model record for "allenai/olmOCR-2-7B-1025-FP8", the olmOCR-2 weights the README's examples run, tagged "license:apache-2.0".

• https://raw.githubusercontent.com/allenai/olmocr/main/LICENSE recorded 2026-09-28

  The raw LICENSE is the full Apache License text, "Apache License Version 2.0, January 2004".

• https://raw.githubusercontent.com/allenai/olmocr/main/README.md recorded 2026-09-28

  The raw README installs the pipeline with "pip install olmocr", points at training code ("SFT Finetuning code for Qwen2.5-VL"), runs the model locally with "vllm serve allenai/olmOCR-2-7B-1025-FP8" and states the project "is licensed under [Apache 2.0]". Searched the body for enterprise, subscription and license key: none.

Adoption is measured as PyPI downloads of the olmocr package over the trailing month. That counts direct installs of the library, not use of olmOCR inside the Dolma 3 pipeline it also feeds.

• https://pypistats.org/api/packages/olmocr/recent recorded 2026-08-13

  22,925 downloads in the trailing 30 days for olmocr

olmOCR's pipeline turns PDFs and images into text with reading order, tables and equations recovered, but it does not emit typed entities a customer defines. As corroboration, it scores 82.4 on its own olmOCR-Bench, ahead of MinerU's 75.2 but behind Chandra's 83.1, and 85.74 on OpenDataLab's OmniDocBench, where it sits in the lower half of a board led by TeleOCR at 96.91.

• https://raw.githubusercontent.com/allenai/olmocr/main/README.md recorded 2026-09-28

  The raw README lists "Support for equations, tables, handwriting, and complex formatting" and "Automatically removes headers and footers", and its olmOCR-Bench table gives "Chandra OCR 0.1.0*" at "83.1±0.9", "Infinity-Parser 7B*" at "82.5±?", olmOCR v0.4.0 at "82.4±1.1", and Marker, MinerU and Mistral OCR at "76.1±1.1", "75.2±1.1" and "72.0±1.1".

• https://raw.githubusercontent.com/opendatalab/OmniDocBench/main/README.md recorded 2026-09-28

  The OmniDocBench README table "Comprehensive evaluation of document parsing on OmniDocBench (v1.6_full)" is led by "TeleOCR" at "96.91", has "MinerU2.5-Pro" at "95.75" and "MinerU-2.5" at "93.04", and lists "olmOCR" at "85.74", below GPT-5.2 at "86.59" and above "Mistral OCR" at "85.66".