SoMarkAI/OXR

An open-source document parsing model with SOTA accuracy and speed, using an RSI loop during post-training.

What it solves

Documents like PDFs and scanned images are hard for AI agents to read directly. OXR turns them into structured, machine-friendly data (Markdown or JSON) while also providing layout coordinates, so downstream tools know where each piece of content sits on the page. This saves developers from building custom parsing pipelines and makes document data usable in automated workflows.

How it works

OXR is a compact 1.3B-parameter model trained from scratch for document parsing. It uses two separate components: a layout analysis model that identifies regions (tables, figures, text, etc.) and the OXR recognition model that extracts content and determines reading order. This split design lets you swap in your own layout model if needed. During post-training, OXR uses a recursive self-improvement (RSI) loop — it automatically targets its own weak or wrong outputs to improve without human annotation.

The system outputs content for 13 element types (including tables, formulas, chemical structures, code blocks) with bounding boxes. It supports challenging layouts like A3 exam papers, handwritten notes, and vertical text.

Who it’s for

Developers and teams who need to convert documents into structured data for AI agents, RAG pipelines, or other automation. It suits anyone who wants a fast, accurate, open-source parser that can be run locally or accessed via a free API (for testing) and customized for domain-specific layout needs.

Highlights

  • #1 ranking on three document parsing benchmarks (olmOCR-Bench, OmniDocBench, DocParsingBench) with a particularly strong lead in table parsing.
  • Pluggable layout analysis — the built-in layout model can be replaced without losing recognition and reading order capabilities.
  • Speed-focused design with only 1.3B parameters, achieving top-tier parsing speed on an RTX PRO 6000.
  • Free API for quick testing without signup (not recommended for production use).

Related

  • Project
    jd-opensource/JoyAI-Image

    A unified multimodal foundation model that integrates image understanding, text-to-image generation, and instruction-guided image editing with a focus on spatial intelligence.

  • Project
    MiniMax-AI/cli

    The official command-line interface for the MiniMax AI platform, enabling direct terminal access to text, image, video, and speech generation models and simplifying agent configuration.

  • Project
    zai-org/GLM-V

    A series of vision-language models (VLMs) that enhance multimodal reasoning, supporting native function calling, visual grounding, and complex document understanding.

  • Project
    OpenGVLab/InternVideo

    A series of video foundation models and large-scale datasets designed for multimodal video understanding, long-context modeling, and contextual reasoning.