The document parsing field has just welcomed another formidable new contender.
OXR is officially open-sourced: its main model has only 1.3B parameters, topping all three benchmarks—OmniDocBench, olmOCR-Bench, and DocParsingBench—with complete parsing speed surpassing the also-competing PaddleOCR-VL-1.6 and MinerU2.5-Pro, pushing both the accuracy and efficiency of lightweight models to the top of the charts.The SoMark team behind it comes from a domestic startup deeply engaged in the document intelligence field, which has continuously refined its parsing capabilities around its core product SoMark, and now introduces the open-source model OXR.
The full name of OXR is Optical Everything Recognition. From OCR to OXR, what OXR must handle goes far beyond text. It can parse 13 types of document elements including tables, formulas, and chemical expressions, accurately restore the reading order of multi-column layouts, output Markdown / JSON, and preserve layout coordinates, turning complex documents into data that Agents can continue to use.
1 High-precision parsing, first place on all three global leaderboards
According to the evaluation results released by the team,OXR scored 97.18, 87.84, and 87.76 on OmniDocBench, olmOCR-Bench, and DocParsingBench respectively, ranking first in all of them.
Beyond the scores, an even more intuitive look is how OXR handles complex documents.
For a table spanning multiple rows and columns, correct text recognition is only the first step. Once row-column relationships or merged cells are misaligned, numbers may be mapped to the wrong items. OXR can preserve these structures and output tables as HTML for subsequent display and processing.
The challenges in scientific research documents are even denser: mathematical formulas have superscripts, subscripts, and nested structures, while chemistry content includes reaction equations and molecular structure diagrams. OXR outputs formulas in LaTeX, represents chemical equations with mhchem, and represents chemical structural formulas with SMILES, turning this content from graphics on a page into expressions that can be further processed.
Multi-column papers, newspapers, and exam papers also test reading order. Even if every region is recognized correctly, cross-reading of left and right columns, or mixing body text with marginal notes, will still destroy the original meaning. OXR determines the reading order while recognizing content, and provides layout coordinates, preserving the position of content on the page.
These capabilities together constitute OXR's "Everything": covering 13 types of document elements including text, tables, formulas, chemical structural formulas, and code blocks, parsing PDFs and document images into Markdown / JSON for downstream processes such as Agents and RAG.
2 Dual SOTA in accuracy and speed: not just accurate, but fast
When processing a single document, a difference of a few seconds may not be noticeable; when the task becomes batch paper ingestion, archive organization, or research report parsing, processing efficiency directly affects waiting time and compute usage.
According to the team's released complete parsing test results under identical conditions, on the same RTX PRO 6000 Blackwell (96GB) GPU, parsing 100 pages of documents randomly sampled from OmniDocBench,OXR took an average of about 101 seconds, while PaddleOCR-VL-1.6 and MinerU2.5-Pro needed about 145 seconds and 255 seconds respectively.
Small parameter count, high accuracy, and fast speed—all appeared in the same model this time.
This is thanks to a clever design by OXR in task division:a lightweight layout analysis model locates content regions, and then the main model completes content recognition and reading order determination.
Merging ordering into the content recognition stage reduces the additional overhead of separately handling reading order, and also allows the model to judge sequence relationships based on content semantics. This is an architectural design of OXR that balances parsing quality and efficiency.
3 RSI: continuous improvement around erroneous and weak samples
Behind the impressive results, OXR's training approach is equally worth examining. As we all know, document parsing can hardly cover all problems with a fixed set of samples. Scenarios, languages, and layouts keep changing—how can a model stably output reliable parsing results even when facing complex documents it has never seen?
OXR innovativelyintroduces for the first time in the post-training stage a Recursive Self-Improvement (RSI) training paradigm oriented toward document parsing,converting erroneous samples and weak samples from parsing into the basis for the next round of training.
This workflow can be understood as an 'error-finding and remedial lesson' process completed through the collaboration of multiple AI models:
- OXR first parses unlabeled documents, producing results to be evaluated.
- External models judge the results and analyze the causes of errors sample by sample, while an independent clustering model categorizes the problems by cause.
- External models then generate targeted training data around these problems, to be used for the next round of OXR training.
In this way, the preparation of training data gains a clear problem orientation: wherever the model is weak, data is supplemented around that area, accelerating the pace of model iteration. Meanwhile,the training data added in this closed loop requires no manual annotation,also breaking through the efficiency bottleneck caused by manually prepared materials.
4 Pluggable layout analysis for more flexible business adaptation
When a model misreads, accuracy needs to be improved; but in many cases, parsing results do not meet expectations because the business defines document structure differently.
For example, with the same set of images, archive management may want to keep them as a whole, while image retrieval requires extracting them one by one; for a region containing a title, text, and a border, some businesses split it as text, while others want it treated as a table as a whole. Both requirements may be reasonable.
To address this, OXR designs layout analysis as an independent, pluggable module.
You can use the built-in layout analysis model directly, or integrate other layout analysis models that conform to your own business standards, adjusting element categories, region boundaries, and content grouping, and then continue to reuse OXR's content recognition and reading order determination capabilities.
For developers building RAG, Agents, and document workflows, OXR undoubtedly offers an open-source option that balances parsing accuracy, processing speed, and layout adaptability.
Currently, OXR has released its code and model weights, and provides an online Demo and free API so developers can test its performance directly with real documents:
- GitHub repository: https://github.com/SoMarkAI/OXR
- Hugging Face model weights: https://huggingface.co/SoMarkAI/OXR-1.0
- Online Demo: https://free-oxr.somark.cn/
- Free API: https://free-oxr.somark.cn/api/v1
Teams with more production-grade document parsing needs can also learn about SoMark, a document parsing engine built by the same team and powered by an independent model.
SoMark official website: https://somark.cn/
From SoMark to OXR, this team is opening up the valuable experience they have accumulated through long-term exploration and practice to developers, conveying their values: turning the knowledge in documents into high-quality data usable by AI, and bringing it into more real business scenarios.
This article was provided by the SoMark team and republished by QbitAI with authorization; the views belong to the original authors.