What is QwenLong-L1.5
QwenLong-L1.5 is a large language model focused on long-context reasoning, launched by the Alibaba Tongyi Lab. The model is based on the Qwen3-30B-A3B architecture and, through a systematic post-training scheme, significantly improves long-context reasoning capabilities. Through a high-quality data synthesis pipeline, a stable reinforcement learning method, and a memory management framework that breaks through physical window limitations, the model addresses the shortcomings of traditional models in long-context tasks, such as training instability and limited context windows. The model performs excellently on multiple long-context reasoning benchmarks, with performance approaching or even surpassing GPT-5 and other top models, while also achieving significant improvements in general capabilities.
Main Features of QwenLong-L1.5
-
Long-Context Reasoning: The model can process and reason over ultra-long texts (such as documents with millions of tokens), solving complex problems that require cross-paragraph and cross-document understanding.
-
Multi-Hop Reasoning: Supports multi-hop logical reasoning, that is, chaining information through multiple steps to complete complex reasoning tasks.
-
Information Integration: Extracts and integrates scattered information from long texts to answer questions that require global understanding.
-
Memory Management: Through a memory enhancement framework, it handles tasks that exceed the physical context window, enabling efficient reasoning over ultra-long texts.
-
Improved General Capabilities: It shows significant improvements on general tasks such as mathematical reasoning, agent memory, and long conversations, demonstrating good generalization ability.
Technical Principles of QwenLong-L1.5
- High-Quality Data Synthesis Pipeline: QwenLong-L1.5 builds high-quality long-context reasoning data through a "decompose first, then compose" approach. The model decomposes 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 Method: QwenLong-L1.5 introduces task-balanced sampling and task-specific advantage estimation strategies to address data distribution shift and unstable reward signals in multi-task long-context training. At the same time, it proposes the Adaptive Entropy-controlled Policy Optimization (AEPO) algorithm, which dynamically regulates negative gradients to balance exploration and exploitation in the model, effectively solving instability in long-context reinforcement learning and ensuring stable training on longer sequences.
- Memory Management Framework That Breaks Through the Physical Window: A memory management framework is designed that, through multi-stage fusion reinforcement learning training, combines single-pass inference with iterative memory processing. When handling ultra-long texts, the model does not rely on a limited context window. Through chunked processing and iterative memory updates, it "folds" global information into a compact memory representation, breaking through the limitations of the physical window and enabling efficient reasoning over ultra-long texts.
Project Address of QwenLong-L1.5
- GitHub Repository：https://github.com/Tongyi-Zhiwen/Qwen-Doc
- HuggingFace Model Library：https://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 Analysis: Used to analyze financial reports, legal documents, and academic literature, extract key information, and perform multi-hop reasoning, helping professionals quickly understand complex content.
-
Code Understanding and Generation: Helps developers understand large codebases, generate code snippets and completion suggestions, and improve programming efficiency and code quality.
-
Complex Question-Answering Systems: Handle multi-hop question answering and long-document queries, providing accurate, in-depth answers for intelligent customer service and knowledge question-answering systems.
-
Information Retrieval and Integration: Optimize search engine results, extract knowledge from long texts to build knowledge graphs, and improve the accuracy and efficiency of information retrieval.
-
Education and Learning: Assist online education platforms and intelligent tutoring systems, help students understand complex academic problems, and support personalized learning.