SoMark released OXR, an open-weight document parser for AI agents

A coding agent that needs to read a PDF usually sends the file to a hosted service. That upload is the step that makes a client contract or a signed form hard to touch.

The SoMark team put OXR on GitHub with 842 stars. The repo is 21 days old. OXR is a 1.3B-parameter model. It turns a PDF or a scanned page into Markdown or JSON. Each block comes back with layout coordinates. The weights are on Hugging Face. The licence is called "OXR Open Source License, a custom license based on Apache 2.0 with additional terms". A buyer should read it before putting this into a product.

What it recognises

The repo says OXR covers 13 element types in one pass. Plain text, tables, formulas, code blocks, chemical equations and handwriting are on the list. OXR returns the reading order with the content. A long two-column PDF comes back as one Markdown file in the right order. The project includes two models together. One is a layout analysis step. The other is the OXR recognition model. It handles the content and the reading order in one pass.

The claims and who makes them

SoMark says OXR is "#1 across three benchmarks": olmOCR-Bench, OmniDocBench, and DocParsingBench. These are the team's own numbers against the public benchmark sets. No third party has run the test and published a result. A buyer who cares which parser is best on their own documents should run the test themselves.

The team also describes an "RSI loop" used during post-training. It is a recursive self-improvement pass. The model keeps picking the samples it gets wrong and training on them, without a human relabelling. The claim is that this pushed the final numbers up without more labelled data. That is also the team's own description of its own training pipeline.

Who it is for

A small team that wants an agent to read a PDF. The team does not want the file to leave the laptop. OXR is under 2B parameters. It fits on one GPU. A developer can test it on a workstation. A team already paying a hosted API for document parsing has a reason to measure the two. Run both on your own files and compare.

The limits are honest to name. The licence is not plain Apache 2.0. A company that resells output or trains on top should read it first. The benchmarks are the vendor's own. The project is 21 days old. The people running it in production are a small number. A careful team should test it with care and not use it as a default yet.

Source

Primary source: SoMarkAI/OXR on GitHub. Weights: SoMarkAI/OXR-1.0 on Hugging Face.