SoMark open-sources OXR-1.0, a 1.3B document parser claiming first place on three benchmarks

SoMark, a Chinese startup focused on document intelligence, has open-sourced OXR (Optical Everything Recognition), a document parsing model whose main model has 1.3B parameters. According to the team's published results, OXR ranks first on OmniDocBench, olmOCR-Bench and DocParsingBench. The report originates from the SoMark team and was republished with authorization by Chinese outlet QbitAI (量子位); the original is in Chinese.

Benchmarks and speed

The team reports these scores, all first place on their respective benchmarks:

Benchmark	OXR score
OmniDocBench	97.18
olmOCR-Bench	87.84
DocParsingBench	87.76

For speed, the team ran a same-conditions full-parsing test on one RTX PRO 6000 Blackwell (96GB) GPU, using 100 pages randomly sampled from OmniDocBench. OXR averaged about 101 seconds. PaddleOCR-VL-1.6 needed about 145 seconds and MinerU2.5-Pro about 255 seconds. These are the team's own measurements.

According to QbitAI, the repository passed 1,000 GitHub stars in its first week of open-source availability, and attention on X exceeded one million.

What OXR parses

OXR handles 13 document element types, including text, tables, formulas, chemical structures and code blocks. It converts PDFs and document images into Markdown or JSON and keeps layout coordinates. Output formats by element:

• Tables: HTML, preserving merged cells and row/column relationships
• Formulas: LaTeX
• Chemical equations: mhchem
• Chemical structures: SMILES

The model also determines reading order for multi-column papers, newspapers and exam papers.

Architecture

A lightweight layout analysis model locates content regions. The main 1.3B model then performs content recognition and reading-order judgment in a single stage. The team says merging ordering into recognition removes the overhead of a separate reading-order step and lets the model use content semantics to decide sequence.

The layout module is pluggable. Developers can use the built-in layout model or substitute their own to change element categories, region boundaries and content grouping, while reusing OXR's recognition and reading-order capabilities.

Training: recursive self-improvement

The team says OXR is the first to apply a Recursive Self-Improvement (RSI) paradigm for document parsing in the post-training stage. The loop works as follows:

1. OXR parses unlabeled documents.
2. An external model judges the results and analyzes errors per sample, and an independent clustering model groups the problems by cause.
3. An external model generates targeted training data for those problem categories, which feeds the next training round.

The team states that the added training data in this loop needs no manual annotation.

Availability

Code and weights are open. SoMark also offers an online demo and a free API.

• GitHub: github.com/SoMarkAI/OXR
• Weights: huggingface.co/SoMarkAI/OXR-1.0
• Demo: free-oxr.somark.cn
• Free API: free-oxr.somark.cn/api/v1

SoMark separately sells a production document parsing engine, also called SoMark, which is driven by an independent model.

What this means

All benchmark and speed figures come from the SoMark team, so they are claims until independently reproduced. The speed comparison covers one GPU and one 100-page sample from OmniDocBench. If the results hold, a 1.3B model that parses faster than the two named baselines gives teams building RAG and agent pipelines a smaller option to run on their own hardware. The pluggable layout module addresses cases where a business defines document structure differently from the default, such as keeping an image set whole versus extracting images individually. Open weights and a free API let developers test OXR on their own documents.