QwenLong-L1.5 - A long-text inference model open-sourced by Alibaba Tongyi

What is QwenLong-L1.5?

QwenLong-L1.5 is a large language model developed by Alibaba's Tongyi Lab, focusing on long text reasoning. Based on the Qwen3-30B-A3B architecture, the model significantly improves long text reasoning capabilities through a systematic post-training scheme. By employing a high-quality data synthesis pipeline, stable reinforcement learning methods, and a memory management framework that overcomes physical window limitations, the model addresses the shortcomings of traditional models in long text tasks, such as unstable training and limited context windows. The model performs exceptionally well in multiple long text reasoning benchmarks, achieving performance close to or even surpassing top models like GPT-5, while also showing significant improvements in general-purpose capabilities.

Main functions of QwenLong-L1.5

• Long text reasoningThe model can process and reason about extremely long texts (such as documents with millions of tokens), solving complex problems that require cross-paragraph and cross-document communication.
• Multi-hop reasoningIt supports multi-hop logical reasoning, that is, it connects information through multiple steps to complete complex reasoning tasks.
• Information integrationExtracting and integrating scattered information from long texts to answer questions that require a global understanding.
• Memory Management: By using a memory-enhanced framework, it can handle tasks that go beyond the physical context window, enabling efficient reasoning on extremely long texts.
• General ability enhancementIt shows significant improvements in general tasks such as mathematical reasoning, agent memory, and long dialogues, demonstrating good generalization ability.

Technical Principles of QwenLong-L1.5

• High-quality data synthesis pipelineQwenLong-L1.5 constructs high-quality long-text reasoning data through a "deconstruction-then-recombination" approach. The model breaks down long documents into atomic facts and their relationships, and uses tools such as knowledge graphs and multi-document tables to programmatically synthesize complex problems that require multi-hop reasoning and global information integration.
• Stable reinforcement learning methodsQwenLong-L1.5 introduces task-balanced sampling and task-specific advantage estimation strategies to address data distribution bias and reward signal instability in multi-task training of long texts. Simultaneously, it proposes an Adaptive Entropy Control Optimization (AEPO) algorithm, which dynamically adjusts the negative gradient to balance model exploration and utilization, effectively solving the instability in long text reinforcement learning and ensuring stable training of the model on longer sequences.
• Breaking through the memory management framework of physical windowsThe design employs a memory management framework that combines single-inference with iterative memory processing through multi-stage fusion reinforcement learning training. When processing extremely long texts, the model does not rely on a limited context window. Instead, it "folds" global information into a compact memory representation through block processing and iterative memory updates, overcoming the limitations of physical windows and achieving efficient reasoning for extremely long texts.

QwenLong-L1.5 project address

• GitHub repository: https://github.com/Tongyi-Zhiwen/Qwen-Doc
• HuggingFace model libraryhttps://huggingface.co/Tongyi-Zhiwen/QwenLong-L1.5-30B-A3B
• arXiv technical paper: https://arxiv.org/pdf/2512.12967

Application Scenarios of QwenLong-L1.5

• Long document analysisIt is used to analyze financial reports, legal documents, and academic literature, extract key information, and perform multi-hop reasoning to help professionals quickly understand complex content.
• Code understanding and generationIt helps developers understand large codebases, generates code snippets and completion suggestions, and improves programming efficiency and code quality.
• Complex question answering systemIt handles multi-hop question-and-answer and long document queries, providing accurate and in-depth answers for intelligent customer service and knowledge-based question-and-answer systems.
• Information retrieval and integrationOptimize search engine results, extract knowledge from long texts to build knowledge graphs, and improve the accuracy and efficiency of information retrieval.
• Education and LearningIt assists online education platforms and intelligent tutoring systems to help students understand complex academic problems and supports personalized learning.