SoMark's 1.3B document parsing model, pre-trained from scratch, converts PDFs and document images into Markdown or JSON with layout coordinates; layout analysis is handled by an independent, replaceable small model, while OXR itself is responsible for content recognition and reading order, capable of identifying 13 types of elements including tables, formulas, seals, code blocks, chemical equations (mhchem), and chemical structures (SMILES).
Score increase 84 · Likes 125 · Downloads 2481
Reasons
- Official GitHub evaluation documentation (snapshot from 10 5): olmOCR-bench 87.84 (second highest is Infinity-Parser2-Pro at 87.60), OmniDocBench 97.18 (TeleOCR 96.91, PaddleOCR-VL-1.6 96.34), DocParsingBench 87.76 (PaddleOCR-VL-1.6 85.17), ranking first in all three benchmarks.
- Officially claimed to achieve serial decoding of 557.77 tokens/second and end-to-end processing of approximately 0.99 pages/second on a single RTX PRO 6000 Blackwell GPU, making it the fastest among the 21 compared models; post-training employed a Recursive Self-Improvement (RSI) process that automatically iterates on errors, requiring no manual annotation.
- Offers a registration-free free API and online demo, with an interface compatible with SoMark's official CLI/SDK; yesterday it lost points because results were only shown in images, but this time it is included with textual numerical data from the evaluation documentation, which was not covered in previous editions.
Usage
- Try first: Online demo at free-oxr.somark.ai (domestic mirror somark.cn); after
pip install somark, runsomark --base-url https://free-oxr.somark.ai/api/v1 --api-key dummy parse file.pdf --formats md,json. - Local (Docker recommended):
git clone SoMarkAI/OXR,hf download SoMarkAI/OXR-1.0 --local-dir ./models/OXR-1.0, then build and run usingdocker build -f docker/Dockerfile.allinone, and openlocalhost:8000/demo; for source code deployment, start three separate processes: the OXR model service (vLLM 0.25.1 + repository-provided plugin), the layout analysis service, and the parsing API. - Parsing interface:
curl localhost:8000/v1/parse/sync -F [email protected] -F output\_formats=markdown -F output\_formats=json.
Risks
- Requires Linux + NVIDIA GPU supporting BF16 (CUDA 13.0 driver); minimum 24GB VRAM for single-GPU setup, with 48GB or more recommended for long documents or concurrency. Official vLLM does not yet natively support it, requiring the out-of-tree plugin from the repository.
- The license is the OXR Open Source License, which adds additional terms to Apache 2.0; review the terms before commercial use; the technical report has not yet been released.
- All scores are officially aggregated, with scores for compared models taken from their respective papers or leaderboards; there is no direct comparison with MiniCPM-OCR 1.0, which is also 1.3B; the free API is officially stated to not guarantee stability and is not recommended for production use.