Qwen-Doc

An Open-Source Collection of Projects on Document Understanding, Parsing, and Agents

📖 Introduction

Qwen-Doc is an open-source repository dedicated to Document AI, developed and maintained by the Tongyi-Zhiwen team.

This repository aims to bring together a series of explorations and practices centered on cutting-edge technologies such as long-context understanding, document parsing, and document-based intelligent agents. We are committed to enhancing the capabilities of Large Language Models in processing and comprehending complex documents, and we open-source our models, data, and methodologies to foster community growth.

🎉 News

• May 21, 2026: 🔥 We released the MPDocBench-Parse, a practical benchmark for multi-page document parsing based on 420 real-world PDFs (3,135 images) from 15 domains. Beyond text/formula/table fidelity, it also evaluates semantic continuity, figure extraction, and document-level logical structure.
• Dec 15, 2025: 🔥 We released the QwenLong-L1.5 project! It provides a complete post-training recipe for long-context reasoning and memory management. The corresponding model and technical report have also been released.
• Dec 15, 2025: 🔥 We released the code implementation of SPELL, which is a self-play reinforcement learning framework designed to improve long-context reasoning abilities in LLMs.
• May 28, 2025: 🔥 The QwenLong-L1 project released QwenLong-L1-32B-AWQ, a version processed with AWQ int4 quantization.
• May 26, 2025: 🔥 We officially open-sourced the QwenLong-L1 project, the industry's first large model trained for long-context reasoning using reinforcement learning. We also released the accompanying QwenLong-L1-32B model and the DocQA-RL-1.6K training dataset.

📂 Project List

This repository currently includes the following projects:

1. QwenLong-L1

• Description: A framework designed to generalize Large Models from short-context proficiency to robust long-context reasoning capabilities using Reinforcement Learning. This project explores mechanisms like curriculum learning and difficulty-aware sampling, and releases the QwenLong-L1-32B model trained on this framework, which has achieved state-of-the-art performance on multiple long-context document question answering (DocQA) benchmarks.

2. QwenLong-L1.5

• Description: A complete "Post-Training Recipe" for long-context reasoning and memory management. This project features three core contributions: a synthesis pipeline for generating complex reasoning data, the Adaptive Entropy-Controlled Policy Optimization (AEPO) algorithm optimized for long-context training, and a memory management framework that extends operation beyond the model's physical context window. Based on this recipe, we introduce the QwenLong-L1.5-30B-A3B model.

3. SPELL

• Description: A self-play reinforcement learning framework designed to improve long-context reasoning abilities in LLMs. SPELL cycles a single LLM through three roles—questioner, responder, and verifier—to autonomously generate training data and rewards, without requiring external supervision. Extensive experiments across 12 models and 6 benchmarks demonstrate consistent improvements. Notably, SPELL provides a potential path for elevating the performance ceiling of models surpassing human performance.

4. MPDocBench-Parse

• Description: MPDocBench-Parse is a comprehensive benchmark and evaluation toolkit designed for practical multi-page document parsing. It is constructed from 420 real-world PDF documents, comprising 3,135 page images, and covers 15 domains across both English and Chinese. In addition to assessing element-level fidelity, including text, formulas, and tables, MPDocBench-Parse further incorporates several critical yet underexplored evaluation dimensions in existing benchmarks: semantic continuity, figure extraction, and document-level logical structure.

⭐ Star History

📝 Citation

If you find our work helpful in your research, please consider citing our papers:

@article{wan2025qwenlongl1,
  title={QwenLong-L1: : Towards Long-Context Large Reasoning Models with Reinforcement Learning},
  author={Fanqi Wan, Weizhou Shen, Shengyi Liao, Yingcheng Shi, Chenliang Li, Ziyi Yang, Ji Zhang, Fei Huang, Jingren Zhou, Ming Yan},
  journal={arXiv preprint arXiv:2505.17667},
  year={2025}
}

@article{shen2025qwenlongl15,
      title={QwenLong-L1.5: Post-Training Recipe for Long-Context Reasoning and Memory Management}, 
      author={Weizhou Shen and Ziyi Yang and Chenliang Li and Zhiyuan Lu and Miao Peng and Huashan Sun and Yingcheng Shi and Shengyi Liao and Shaopeng Lai and Bo Zhang and Dayiheng Liu and Fei Huang and Jingren Zhou and Ming Yan},
      journal={arXiv preprint arXiv:2512.12967},
      year={2025}
}

@article{yang2025spell,
    title={SPELL: Self-Play Reinforcement Learning for evolving Long-Context Language Models},
    author={Ziyi Yang, Weizhou Shen, Ruijun Chen, Chenliang Li, Fanqi Wan, Ming Yan, Xiaojun Quan, Fei Huang},
    journal={arXiv preprint arXiv:2509.23863},
    year={2025}
}

@misc{zhou2026mpdocbenchparsebenchmarkingpracticalmultipage,
      title={MPDocBench-Parse: Benchmarking Practical Multi-page Document Parsing}, 
      author={Bangbang Zhou and Hangdi Xing and Yifan Chen and Jianjun Xu and Qi Zheng and Feiyu Gao and Zhibo Yang and Shuai Bai and Ming Yan and Jieping Ye and Hongtao Xie},
      year={2026},
      eprint={2605.22100},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2605.22100}, 
}