Qwen-Doc is an open-source repository focused on Document AI, developed by the Tongyi-Zhiwen team. The repository includes various projects aimed at enhancing document understanding, parsing, and the development of intelligent agents. It emphasizes improving Large Language Models (LLMs) for complex document processing. Key components include QwenLong-L1, a model that transitions from short-context to long-context reasoning using reinforcement learning. This model has set new benchmarks in long-context document question answering. Recent releases showcase significant advancements. On December 15, 2025, QwenLong-L1.5 was launched, providing a comprehensive post-training recipe for managing long-context reasoning and memory. This includes innovative techniques like the Adaptive Entropy-Controlled Policy Optimization (AEPO) algorithm and a memory management framework that extends beyond the model's physical context. The project also produced the QwenLong-L1.5-30B-A3B model. Another notable project, SPELL, enables self-play reinforcement learning where a single LLM assumes different roles to generate training data autonomously. This method has shown consistent improvements across multiple models and benchmarks. The repository's earlier work includes the QwenLong-L1 project, which was the first large model trained for long-context reasoning using reinforcement learning. The QwenLong-L1-32B model was released alongside the DocQA-RL-1.6K training dataset, focusing on curriculum learning and difficulty-aware sampling. These projects collectively aim to push the boundaries of what LLMs can achieve in document understanding and reasoning, making them valuable resources for researchers in the field.