A 1.3B model that out-parses the big OCR guns

What it does

OXR (“Optical Everything Recognition”) is an open-source document parsing model from the SoMark team. It converts PDFs and document images into Markdown or JSON with layout coordinates attached, recognizing 13 element types — tables, formulas, code blocks, stamps, even chemical structures output as SMILES. The pitch is “agent-friendly data”: output you can feed straight to an LLM pipeline instead of wrangling raw OCR text.

The interesting bit

Two things stand out. First, the layout analysis module is deliberately pluggable — the team observed that layout rules vary more across domains than anything else, so you can swap in your own layout model while OXR keeps handling recognition and reading order. Second, post-training uses a recursive self-improvement (RSI) loop that targets samples the model gets wrong, iterating without human annotation — which they claim is a first for document parsing.

Key highlights

• Claims #1 on olmOCR-Bench, OmniDocBench, and DocParsingBench, with a wide lead in table parsing
• Only 1.3B parameters; reportedly the fastest open-source document parser in their RTX PRO 6000 speed tests
• Handles messy real-world documents: A3 exam papers, newspapers, vertical text, handwriting
• Free API with no registration or key — explicitly not for production, and kept online “until we can no longer afford to run it”
• Model weights on Hugging Face and ModelScope; arXiv paper marked “coming soon”

Caveats

• Benchmark claims are self-reported; the arXiv paper isn’t out yet, so the RSI loop details rest on the README’s diagrams
• The free API is best-effort: queued requests, no stability guarantee
• Local run appears Linux-only per the README’s requirements section

Verdict

Worth a look if you’re building document ingestion for agents and want a small, fast, swappable-layout parser. If you need a peer-reviewed track record, wait for the paper.