OXR
An open-source document parsing model with SOTA accuracy and speed, using an RSI loop during post-training.
- agent
- llm-tools
- ocr
- rsi
- vlm
- vlm-ocr
- Stars
- 986
- Forks
- 60
- + today
- +5
- Created
- 22d
Ranking data as of October 26, 2026 (UTC).
Star History
Today, hour by hour
Overview
An open-source document parsing model with SOTA accuracy and speed, using an RSI loop during post-training. It ranks #37 on GitTiger, gaining +5 stars on October 26, 2026 (UTC).
The project is written in Python and has 60 forks. It was created 22d ago.
What it is
OXR is an open-source document parsing model from SoMarkAI, written in Python. Its job is the one OCR tools share: turn a page image into structured content. The description pitches it on two axes, accuracy and speed, and says it was trained with an RSI loop during post-training. The supplied fields do not expand that acronym, so this profile will not guess at the method behind it.
Claims and capabilities
The README extract leads with benchmark placements: first on olmOCR-Bench, OmniDocBench and DocParsingBench, with the widest margin said to be in table parsing. It also states that the model recognises thirteen document element types, spanning formulas, tables, chemical equations and code blocks, and that output carries bounding box coordinates for each element. Harder inputs are called out too: oversized A3 exam papers, newspaper pages and vertical text. Those results are the project's own reporting and were not reproduced here.
Getting started
There are no installation or inference commands in the evidence, so the README is where to look for the current procedure. No licence is named in the supplied fields either: the metadata reports no recognised licence, which is not the same as having none, so anyone planning commercial use should clarify the terms.
Missing evidence
No model sizes, latency figures or evaluation methodology appear in the material, so the leaderboard claims should be treated as claims until benchmark details are checked.
git clone https://github.com/SoMarkAI/OXR.git
cd OXR
# see README for setupFrequently Asked Questions
Which benchmarks does OXR report results on?
The README extract names olmOCR-Bench, OmniDocBench and DocParsingBench, and says the model places first on all three with its widest margin in table parsing.
What kinds of document elements can it recognise?
Thirteen element types are claimed, including formulas, tables, chemical equations, chemical structures and code blocks, and the output is said to include bounding box coordinates for each one.
What licence does OXR use?
The repository metadata reports no recognised licence and the evidence does not name one, so the terms must be confirmed with the project before commercial use.
How do I run the model?
The supplied evidence contains no installation or inference instructions, so the README should be consulted for the current setup steps and hardware requirements.