Article · AI translation
One week after going open source, it has gained more than 1,000 stars on GitHub, attracted over 1 million views on X, and appeared in Today’s News recommendations
The document parsing field has a formidable new contender.
OXR is officially open source: its main model has just 1.3B parameters, ranks first on all 3 benchmarks—OmniDocBench, olmOCR-Bench, and DocParsingBench—and is faster at complete parsing than PaddleOCR-VL-1.6 and MinerU2.5-Pro, which were tested alongside it. It has put a lightweight model at the top for both accuracy and efficiency. The SoMark team behind it comes from a domestic startup focused on document intelligence. Having continually refined parsing capabilities around its core product, SoMark, the team has now released the open-source model OXR.
OXR stands for Optical Everything Recognition. Moving from OCR to OXR, it aims to handle far more than text. It can parse 13 types of document elements, including tables, formulas, and chemical formulas; accurately reconstruct the reading order of multicolumn layouts; output Markdown / JSON; and retain layout coordinates, turning complex documents into data that Agents can continue to use.
1 High-accuracy parsing: First place on 3 global benchmarks
According to evaluation results released by the team, OXR scored 97.18, 87.84, and 87.76 on OmniDocBench, olmOCR-Bench, and DocParsingBench, respectively, ranking first on each.
Beyond the scores, a more direct way to assess OXR is to see how it handles complex documents.
For a table with cells spanning rows and columns, recognizing the text correctly is only the first step. If the row and column relationships or merged cells are misplaced, numbers may be assigned to the wrong items. OXR can preserve these structures and output the table as HTML for subsequent display and processing.
Scientific documents present an even denser set of challenges: mathematical formulas have subscripts, superscripts, and nested structures, while chemical content includes reaction equations and molecular structure diagrams. OXR outputs formulas in LaTeX, chemical equations in mhchem, and chemical structures in SMILES, turning graphics on the page into representations that can be processed further.
Multicolumn papers, newspapers, and exam papers also test reading order. Even if every region is recognized correctly, reading across the left and right columns or mixing body text with side notes can still distort the original meaning. OXR determines reading order as it recognizes content and provides layout coordinates to preserve each element’s position on the page.
Together, these capabilities make up the “Everything” in OXR: it covers 13 types of document elements, including text, tables, formulas, chemical structures, and code blocks, and parses PDFs and document images into Markdown / JSON for downstream workflows such as Agents and RAG.
2 State of the art in both accuracy and speed: Accurate and fast
A difference of a few seconds may be hard to notice when processing a single document. But when the task involves ingesting research papers in batches, organizing archives, or parsing research reports, processing efficiency directly affects waiting time and compute use.
According to the team’s results from complete-parsing tests under the same conditions, OXR took an average of about 101 seconds to parse 100 document pages randomly selected from OmniDocBench on the same RTX PRO 6000 Blackwell (96GB) graphics card. PaddleOCR-VL-1.6 and MinerU2.5-Pro took about 145 seconds and 255 seconds, respectively.
A small parameter count, high accuracy, and fast speed have come together in one model.
This is thanks to a clever design in how OXR divides the work: a lightweight layout analysis model locates content regions, then the main model recognizes the content and determines reading order.
Incorporating ordering into the content recognition stage reduces the extra overhead of processing reading order separately and allows the model to use the meaning of the content to determine sequence. This architectural design helps OXR balance parsing quality and efficiency.
3 RSI: Continuous improvement based on incorrect and weak samples
Behind the impressive results, OXR’s training approach also deserves attention. As we all know, it is difficult for a fixed set of samples to cover every challenge in document parsing. With scenarios, languages, and layouts constantly changing, how can a model still reliably parse complex documents it has never seen before?
In a claimed first for document parsing, OXR innovatively introduces a Recursive Self-Improvement (RSI) training paradigm during post-training, turning incorrect and weak parsing samples into a basis for the next round of training.
This process can be understood as “finding mistakes and filling gaps” through the collaboration of multiple AI models:
OXR first parses unlabeled documents, producing results to be evaluated.
External models evaluate the results and analyze the cause of errors for each sample, while an independent clustering model groups the problems by cause.
External models generate training data targeted at these problems for OXR’s next round of training.
This gives training data preparation a clear focus: add data wherever the model is weak, speeding up model iteration. The new training data added within this feedback loop also requires no manual annotation, overcoming the efficiency bottleneck of preparing materials by hand.
4 Pluggable layout analysis for more flexible business adaptation
When a model reads something incorrectly, its accuracy needs improvement. But often, parsing results fail to meet expectations because a business defines document structure differently.
For example, when handling the same group of images, document archiving may call for keeping them together, while image search requires extracting each one separately. For a region containing a title, text, and a border, some businesses split it into text, while others want to treat the whole region as a table. Both requirements may be reasonable.
To address this, OXR makes layout analysis an independent, pluggable module.
You can use the built-in layout analysis model directly or connect another layout analysis model that meets your business standards. You can adjust element categories, region boundaries, and content grouping while continuing to use OXR’s content recognition and reading-order capabilities.
For developers building RAG, Agents, and document workflows, OXR offers an open-source option that combines parsing accuracy, processing speed, and layout adaptability.
OXR’s code and model weights are now available, along with an online demo and a free API, so developers can test its performance directly with real documents:
GitHub code 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 explore SoMark, a document parsing engine built by the same team and powered by a separate model.
SoMark website: https://somark.cn/
From SoMark to OXR, this team is sharing with developers the valuable experience it has gained through long-term exploration and practice. This conveys its guiding principle: turning the knowledge in documents into high-quality data that AI can use across more real-world business scenarios.
This article was provided by the SoMark team and republished by QbitAI with permission. The views are those of the original author.