SoMark's OXR-1.0 Tops 3 Benchmarks Parsing PDFs 15x Faster
SoMark's 1.3B open-source parser tops three benchmarks, runs 15x faster than rivals, and uses a self-improvement loop to patch its own blind spots.
Takeaways
- SoMark open-sourced OXR-1.0, a 1.3B document parsing model with weights on Hugging Face.
- Ranks #1 on olmOCR-Bench (87.84), OmniDocBench (97.18) and DocParsingBench (87.76).
- Parses 100 OmniDocBench pages in 101 seconds on one RTX PRO 6000, 15x faster than Chandra-OCR-2.
- Introduces an RSI self-improvement loop that turns parsing errors into targeted synthetic training data.
- Pluggable layout module lets teams swap the layout model while keeping recognition and reading order.
- Free hosted API at free-oxr.somark.ai, no signup needed; Docker setup on GitHub.
China-based SoMark has released OXR-1.0, an open 1.3-billion-parameter model that converts PDFs and scanned images into structured Markdown or JSON. Its output preserves reading order and includes bounding boxes that map each recognized element to its position on the source page.
SoMark reports leading scores across three document-parsing benchmarks and high throughput on a single NVIDIA GPU. The release also separates layout analysis from recognition and uses a post-training process called recursive self-improvement, or RSI, to generate new training data from recurring errors.
The reported benchmark lead
OXR ranks first in SoMark’s published results for olmOCR-Bench, OmniDocBench and DocParsingBench. Higher scores indicate better parsing accuracy.
| Benchmark | OXR-1.0 | Comparison |
|---|---|---|
| olmOCR-Bench | 87.84 | Chandra-OCR-2: 85.80; jina-ocr-v1: 83.40 |
| OmniDocBench | 97.18 | Reported first place |
| DocParsingBench | 87.76 | Reported first place; benchmark maintained by SoMark |
Table parsing produced OXR’s widest reported accuracy lead. On a single NVIDIA RTX PRO 6000, the model processed 100 pages in 101 seconds, compared with 308 seconds for GLM-OCR. SoMark also reports 15 times the throughput of Chandra-OCR-2 and 2.52 times that of MinerU2.5-Pro under its test conditions.
Layout becomes a replaceable stage
Conventional document parsers often run layout detection, optical character recognition and reading-order reconstruction as sequential stages. OXR combines recognition and ordering in a parallel stage while packaging layout analysis as an independent, pluggable component.