Overview
OXR-1.0 is an open-source document parser from SoMark, a Shanghai team that released both code and weights. The name stands for Optical Everything Recognition and the model carries 1.3 billion parameters. It turns PDFs and document images into Markdown or JSON with layout coordinates and reading order included, aiming at data an agent can consume directly.
Accuracy is the headline. SoMark reports first place on three benchmarks: 87.84 on olmOCR-Bench, 97.18 on OmniDocBench and 87.76 on DocParsingBench, with the widest margin in table parsing. On speed, a 100-page sample from OmniDocBench parsed in about 101 seconds on a single RTX PRO 6000, ahead of the next fastest system.
Two design choices stand out. Layout analysis is a separate, replaceable module, so a team whose documents define structure differently can swap in its own layout model and still reuse OXR's recognition and reading order. Post-training runs a recursive self-improvement loop: an external model scores OXR's output, a clustering model groups the failure causes, and targeted training data is generated for the next round without human labelling.
Key Features
- Thirteen element types: Body text, titles, tables, formulas, figures, stamps, headers and footers, footnotes, code blocks, chemical equations and chemical structures. Tables come out as HTML or Markdown with merged cells preserved, formulas as LaTeX, chemical equations in mhchem notation and structures as SMILES.
- First place on three benchmarks: 87.84 on olmOCR-Bench, 97.18 on OmniDocBench and 87.76 on DocParsingBench by the vendor's own measurement. The evaluation set covers A3 exam papers, newspapers, vertical text and handwriting.
- Swappable layout analysis: The layout model is independent of the main model. When a business wants an image group kept whole or a bordered region treated as a table, it replaces the layout model and keeps OXR for recognition and ordering.
- Reading order inside recognition: Ordering is folded into the recognition pass and produced by the same model, removing a separate sorting stage. Multi-column papers, newspapers and exams therefore get order decided from content semantics.
- Recursive self-improvement: The loop has an external model review unlabelled documents, flags samples where repeated predictions disagree, clusters the likely causes, and generates targeted training data that passes rule checks before the next round. No manual annotation.
- Modest deployment floor: 24GB of VRAM covers low-concurrency testing, with 48GB or more recommended for long documents. Weights ship with an out-of-tree vLLM plugin that works with the official image, plus an all-in-one Docker image.
Use Cases
- Feeding RAG pipelines and agents, converting PDFs into Markdown or JSON with coordinates.
- Financial reports, contracts and invoices where multi-column layout and dense tables must survive parsing.
- Batch conversion of thousands of pages, using the 1.3B footprint to saturate a single GPU.
- Teams with unusual layout definitions, replacing the layout module without touching recognition.
- Digitising exam papers, newspapers, vertical text and handwritten archives, all covered by vendor examples.
- Validating on your own documents through the free API before committing to self-hosting.
Pros
- 1.3B parameters and first place on three document parsing benchmarks, with the widest lead in tables.
- About 101 seconds for a 100-page end-to-end run, ahead of other open-source parsers.
- Pluggable layout analysis keeps customisation contained.
- Output carries layout coordinates and reading order, ready for retrieval and reconstruction.
- Starts at 24GB of VRAM, with a vLLM plugin and Docker image provided.
- A free, no-signup API makes evaluation cost nothing.
Pricing
Code and weights are open under a custom licence based on Apache 2.0. Commercial use is free for organisations whose total funding and annual revenue are both under $2 million; crossing either threshold requires a commercial licence from SoMark, and offering a competing hosted parsing service is not permitted. Self-hosting carries only hardware cost and starts at 24GB of VRAM. A free no-signup API is available for evaluation and explicitly not intended for production, while the team's SoMark engine covers production workloads.
Summary
OXR pushes document parsing back to engineering detail: what happens when a table crosses a page, what decides reading order, and which layer changes when a business defines structure differently. The three benchmark wins are the outcome; pluggable layout and ordering folded into recognition are the reasons. At 1.3B parameters and a 24GB starting point, self-hosting is affordable for small teams. The revenue threshold in the licence is the clause to read before shipping.
Version History
- OXR-1.0 open sourced: a 1.3B parser leading three document benchmarks (2026-10-27): SoMark released OXR-1.0 with weights, a 1.3B document parser that outputs Markdown or JSON with layout coordinates and reading order. It ranks first on olmOCR-Bench at 87.84, OmniDocBench at 97.18 and DocParsingBench at 87.76, and parses a 100-page set in about 101 seconds on one RTX PRO 6000. Thirteen element types with tables as HTML or Markdown, formulas as LaTeX and chemical structures as SMILES. Layout analysis is pluggable, post-training uses a recursive self-improvement loop, and deployment starts at 24GB of VRAM with a vLLM plugin and a free API for evaluation.