Introducing RolmOCR: A Faster, Lighter Open Source Document Model Built on olmOCR

Earlier this year, the Allen Institute for AI released olmOCR, an open-source document OCR model built from a large-scale open dataset and the Qwen2-VL-7B Vision Language Model (VLM). We were excited to see a high-quality, openly available approach to parsing PDFs and other complex documents — and saw an opportunity to utilize newer foundation models and some lightweight optimizations to potentially improve its performance.

The result is RolmOCR, a drop-in alternative to olmOCR that’s faster, uses less memory, and still performs well on a variety of document types. We're releasing it open source for anyone to try out, explore, or build on.

RolmOCR was trained on the same dataset as olmOCR, completely separate from the models we use in production.

Key changes: olmOCR → RolmOCR

We made three notable changes: 

1. New base model: We swapped in a more recent version of the existing model (Qwen2.5-VL-7B) as the foundation.

2. No metadata inputs: Unlike the original, we don’t use metadata extracted from PDFs. This significantly reduces prompt length, which in turn lowers both processing time and VRAM usage — without hurting accuracy in most cases. 

3. Rotation of training data: About 15% of the training data was rotated to enhance robustness to off-angle documents. We otherwise use the same training set. 

Model comparison: Speed vs. performance tradeoffs

Across a variety of test documents, RolmOCR showed either improved or equivalent OCR performance compared to olmOCR — with much faster inference and lower memory consumption. However, in some select cases, RolmOCR did worse. Below are a few examples comparing model outputs.

Example 1: Handwritten note with annotations

Using the same handwritten note example featured on the olmOCR site, RolmOCR produces more accurate results. It correctly captures characters that were previously misread potentially due to corrupted metadata (for instance, "OCLM" is now correctly recognized as "DCLM"). Additionally, it preserves the correct reading order — such as placing "Deepseek Coder" under the appropriate "Data Mixes" section. These details make big differences in terms of downstream parsing and comprehension.

Example 2: Scanned envelope with handwriting and printed text in French and English

In this case, RolmOCR is able to correctly extract most of the information from this low-contrast image but misses some of the smaller text in the bottom left. When testing olmOCR, much of the ordering and some content is missed, and in many cases of running the model, no output is given at all. These results are likely due to a combination of QWen2.5VL's increased capability, and olmOCR's usage on metadata in its training set - there is none available for this particular image.

Example 3: Academic paper written in LaTeX with a borderless table

This is an example where RolmOCR makes a mistake: it completely omits a subtitle found in a table of an academic paper. These elements are successfully extracted by olmOCR, which benefits from structured metadata embedded in the document. This highlights one tradeoff of our metadata-free approach: when metadata is present and accurate, it can provide valuable context that improves extraction — especially for structured fields like headers.

Try RolmOCR

We're releasing it under the Apache 2.0 license for anyone to try out, explore, or build on. You can find instructions, example code, and model details in the README linked here. We’re excited to share RolmOCR with the open source community and hope it’s a useful tool for anyone working with PDFs or complex document layouts.

While RolmOCR is a strong general-purpose option, our own systems support more advanced capabilities — including non-English inputs, layout-aware parsing, and bounding boxes — for teams with more specialized needs.

If you have feedback on the model or want to see how our Reducto models compare, we’d love to hear from you! Feel free to email us at founders@reducto.ai.