SoMark's OXR-1.0 turns PDFs into Markdown and JSON, but its license has a revenue cap

Chinese document AI company SoMark has released OXR-1.0 with open weights, a model that turns PDFs and document images into structured data that AI agents can use. The model's Hugging Face page went up on October 2, and Alibaba's model platform ModelScope highlighted it again this week. Details and setup steps are in the GitHub repository.

What OXR-1.0 is

OXR-1.0 is a document parsing model: it takes a scanned or digital document and extracts its text, tables and other elements in reading order, as Markdown or as JSON with layout coordinates. The company says the model has 1.3 billion parameters, while Hugging Face's file metadata shows 2 billion. It combines a vision encoder with a language model, and the layout analysis part ships as a separate, pluggable model that teams can swap for their own.

Beyond plain text, the model recognizes 13 element types, including titles, text, footnotes, captions, formulas (LaTeX), tables (HTML or Markdown), figures, stamps, code blocks, chemical equations and chemical structures (SMILES). Headers and footers are filtered out by default. Examples include Chinese and English handwriting, exam papers, newspapers and vertical text; no list of supported languages is given.

The company's measurements

According to SoMark's comparison dated October 5, OXR-1.0 ranks first on all three document-parsing benchmarks: olmOCR-Bench, OmniDocBench and DocParsingBench. The company says OXR-1.0 is also the fastest model in a 21-model speed comparison, measured on a 96 GB NVIDIA RTX PRO 6000 Blackwell GPU; end to end it processes roughly one page per second. The company notes that tokenizers differ between models, so token speeds are not exactly comparable. In post-training, the model was trained to improve itself by turning real document-parsing errors into new training signals.

What it takes to run

According to the setup page, the model needs Linux, an NVIDIA GPU with BF16 support and vLLM; 24 GB of GPU memory is suggested for low-concurrency testing and 48 GB or more for production. Because the official vLLM release does not yet support OXR, a separate plugin is used. Running on CPU only or with tools like llama.cpp is not mentioned. The company also offers a free demo and API but advises against using them in production; documents sent there leave your machine.

What to watch in the license

SoMark ships the model under the "OXR Open Source License", which adds terms on top of Apache 2.0 that take precedence in a conflict:

• Revenue and funding cap: Commercial use needs no separate license only while the company and its affiliates stay under $2 million in total funding and $2 million in annual revenue. Companies above either threshold must obtain a separate commercial license from SoMark before starting.
• Attribution: Those offering the model as an online service to third parties must prominently identify SoMark OXR in the interface or documentation.
• No competing API: Offering a hosted API that competes with SoMark's service is prohibited without written permission, regardless of company size.
• Automatic termination: If any of these terms is breached, the license rights end automatically without notice.

The performance and speed figures rest on SoMark's own measurements and have not been independently verified. Although the model is described as trained from scratch, its configuration files reference Google's SigLIP2 and Alibaba's Qwen3. The company's notice also says the code license does not cover the model weights and that weight terms are provided separately, so the terms for using the weights are not entirely clear. The "open source" label therefore means something different from unrestricted open-source licenses, and companies considering it for a product should review the license with their legal team. Follow assessments from independent developers who try the model on their own documents for other perspectives.