olmOCR is an open-source toolkit from the Allen Institute for AI (Ai2) that converts PDFs into clean, naturally ordered plain text for LLM training and retrieval. It has earned more than 19,700 stars on GitHub and is available under the Apache-2.0 license, with the latest updates in March 2026.
Anyone who has tried to feed PDFs to a language model knows the problem: extracted text arrives scrambled, with columns interleaved, tables destroyed, and equations mangled. olmOCR attacks this by using vision-language models to linearize each page the way a human would read it, preserving reading order and representing tables, formulas, and lists in clean text form. The toolkit is engineered for scale, with a pipeline that can process anything from a handful of documents to millions of PDFs using GPU workers and cloud storage.
Quality here is measured, not assumed. Ai2 developed olmOCR alongside evaluation resources for document conversion, and the toolkit’s models were trained on a large, diverse collection of real documents, from scientific papers to forms and historical scans, which is why it copes with layouts that defeat traditional extractors. The pipeline is also cost-conscious: it is designed to run efficiently at scale, making it feasible to convert back catalogs and web-scale PDF collections that would be prohibitively expensive with naive per-page API calls. For many teams building open datasets, olmOCR has become the default first step.
olmOCR is free and open source under the Apache-2.0 license. You only pay for the GPU compute used to run the vision models.
olmOCR is for teams building LLM training corpora, digitizing document archives, or constructing retrieval systems over scientific papers, reports, and other PDFs where faithful text extraction actually matters.