Convert PDFs and images into LLM-ready text
When you need to turn academic papers with mixed formulas and tables or reports with complex layouts into Markdown files, this tool automates the process. It extracts text from image-based documents and organizes it in a natural reading order, making it useful for building large datasets or analyzing document content.
v0.4.0 model update and benchmark results
The previous version, v0.3.0, fixed errors in automatic rotation detection and hallucinations on empty documents, while v0.2.1 used FP8 format by default to increase processing speed and reduce the number of retries per document. olmOCR-Bench is a benchmark consisting of 1,400 documents and over 7,000 test cases, and v0.4.0 recorded a total score of 82.4 points.
Handling complex document structures
This tool converts PDF, PNG, and JPEG documents into Markdown, supporting formulas, tables, handwriting, and complex formatting. It automatically removes headers and footers and maintains a natural reading order even in layouts divided into multiple columns or containing inserted figures.
Installation environment and execution methods
To run on a local GPU, you need an NVIDIA GPU (RTX 4090, L40S, A100, H100, etc.), at least 12GB of GPU memory, and 30GB of disk space. After creating a Python 3.11 environment, you can install it with the pip install olmocr[gpu] command, and it is recommended to install the flashinfer package together for faster inference. If you do not have a local GPU, you can connect to a remote vLLM server and use it with a lightweight installation (pip install olmocr). Using a Docker image resolves model download and environment setup in one step, making it accessible even for beginners.
Large-scale processing and cost efficiency
When processing millions of pages of PDFs, you can perform tasks in parallel across multiple nodes using AWS S3. Once the first worker node sets up the task queue, the remaining nodes automatically pull and process tasks from the same queue. The conversion cost per page is kept under $200 per million pages, reducing the economic burden when building large datasets. It also integrates with external inference providers such as Cirrascale, DeepInfra, and Parasail, allowing you to use the service without directly managing GPU infrastructure.
Considerations before adoption
This tool is released under the Apache 2.0 license, allowing commercial use. However, because it is based on a 7B parameter model, a GPU environment is essential, and you should consider that local execution requires significant hardware resources. While it recorded 82.4 points on the olmOCR-Bench benchmark, showing performance similar to or higher than competing products, it may still have limitations with certain document types (e.g., very old scans). It is recommended to check the conversion quality in advance by testing your document samples with the web demo or a small-scale local test before adoption.
