AnswersPublished19 sources
SoMark OXR-1.0: A 1.3B Document Parser for AI Workflows
SoMark’s 1.3B OXR 1.0 parses PDFs and document images into Markdown or JSON with layout coordinates, aiming to preserve structure for AI workflows.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- SoMark’s 1.3B OXR 1.0 parses PDFs and document images into Markdown or JSON with layout coordinates, aiming to preserve structure for AI workflows.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is SoMark’s open-source OXR-1.0 (Optical Everything Recognition), announced by the Chinese document-intelligence team in October 2026,. Article summary: OXR-1.0 (“Optical Everything Recognition”) is SoMark’s 1.3-billion-parameter document-parsing model, announced in October 2026, that converts PDFs and document images into Markdown or JSON with layout coordinates—not jus. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fak
SoMark’s OXR-1.0 is a 1.3-billion-parameter document parser that converts PDFs and document images into Markdown or JSON with layout coordinates. That structure can make extracted content more useful to retrieval-augmented generation (RAG) and agent pipelines than plain OCR text, but the model’s reported benchmark results should be treated as claims to test against your own documents.4
5
7
What OXR-1.0 does
OXR is designed to extract document content while retaining information about where it appeared on the page. SoMark says it recognizes 13 document-element types; the project materials name examples including tables, formulas, code blocks, stamps, chemical equations and chemical structures, with bounding-box coordinates.4
5
Those outputs may help downstream systems distinguish a table from a paragraph or keep content connected to its location in a source document. The practical value depends on whether the parser preserves the structures and reading order your application needs; benchmark rankings alone cannot establish that for every document type.
How its layout and training approach work
SoMark describes OXR’s layout analysis as independent and pluggable, so a team can substitute a layout model for its own needs. A third-party model description likewise identifies layout analysis as a separate component that can be swapped.8
22 This could be useful for workflows with specialized page layouts, but it is worth testing the chosen layout component alongside recognition and reading-order results.
The project also describes a recursive self-improvement loop during post-training. In SoMark’s account, parsing errors help guide improvement; a secondary description says the process uses an external vision-language model and clusters failure cases.7
8 This is a training approach, not evidence that a deployed model automatically learns from each document it processes.
What the reported benchmark results show
SoMark’s published comparison snapshot lists OXR-1.0 at 97.18 on OmniDocBench, 87.84 on olmOCR-Bench and 87.76 on DocParsingBench, ranking it first among the models in that comparison. The project also highlights a lead in table parsing.1
5
These are benchmark scores, not a universal accuracy percentage or a guarantee of performance on a particular company’s files. Results apply to the models and evaluation protocols in the project’s comparison, so teams should test representative documents—including complex tables and multi-column pages—before relying on OXR in production.1
SoMark’s materials also report high parsing throughput. One comparison snapshot lists 557.77 native tokens per second and 0.990373 pages per second; other reporting describes a 100-page test completed in about 101 seconds.1
31 Treat these as reported results rather than a promise of the same speed on different hardware, document mixes or deployment settings.
Availability and license considerations
The OXR code is available on GitHub, and model weights are listed on Hugging Face.4
13 Before adopting it, review the license carefully: the repository describes Apache License 2.0 as subject to additional terms, including provisions about commercial use.
26 “Open-source” should not be read as unrestricted commercial permission.
Is OXR a fit for a RAG or agent pipeline?
OXR is worth evaluating when a pipeline needs document structure—not just recognized text—and when Markdown or JSON with page coordinates fits the downstream workflow. Its reported scores and speed make it a candidate, not a proven drop-in replacement.1
4
7
Run it on your own sample set and inspect table fidelity, reading order, coordinate quality, latency and the license terms. Those checks will tell you more about suitability for a specific pipeline than a leaderboard position alone.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "SoMark OXR-1.0: A 1.3B Document Parser for AI Workflows"?
SoMark’s 1.3B OXR 1.0 parses PDFs and document images into Markdown or JSON with layout coordinates, aiming to preserve structure for AI workflows.
Sources
github.comGitHub - SoMarkAI/OXR
somark.aiErrors as Data: How RSI Helps OXR Achieve SOTA Accuracy and Speed in Document Parsing
fourweekmba.comOpen 1.3B OXR Tops 3 Parsing Tests, With $2M Licence ...
slm.expertOXR-1.0 — slm.expert