A new 1.3 billion parameter open-source OCR model named OXR has achieved top rankings on three global benchmarks for document parsing. Developed in China, OXR can accurately extract various document elements including text, tables, formulas, and images, and converts them into structured formats like Markdown or LaTeX. The model's efficiency is highlighted by its ability to process 100 pages in just over 100 seconds on a single high-end GPU, making it accessible for local deployment. OXR's performance is attributed to a novel Recursive Self-Improvement (RSI) training paradigm, which allows the model to learn from its own errors and generate new training data without human annotation. AI
IMPACT This model's efficiency and accessibility could significantly accelerate enterprise adoption of advanced document processing and RAG systems.
RANK_REASON New open-source model release with benchmark performance claims. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- DocParsingBench
- GeForce RTX 3060
- General VLM
- GitHub
- LaTeX
- Nvidia RTX Pro 6000 Blackwell Workstation Edition
- olmOCR-Bench
- OmniDocBench
- SigLIP2
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
