What is QwenLong-L1-32B
QwenLong-L1-32B is the first open-source long-textreasoning large modelfrom the Qwen-Doc team at Alibaba Group. The model is based on progressive context extension, curriculum-guided reinforcement learning, and a difficulty-aware retrospective sampling strategy, significantly improving reasoning capabilities in long-text scenarios. The model performs excellently on multiple long-text document question answering (DocQA) benchmarks, achieving an average accuracy of 70.7%, surpassingOpenAI-o3-miniand existing flagship models such as Qwen3-235B-A22B, and is comparable to Claude-3.7-Sonnet-Thinking. QwenLong-L1-32B can handle complex multi-hop reasoning, logical reasoning, and mathematical reasoning problems, and is suitable for multiple fields such as law, finance, and scientific research, demonstrating powerful long-text processing and reasoning capabilities.
Main features of QwenLong-L1-32B
- Long-text reasoning: Handles complex long-text tasks, such as multi-hop reasoning, logical reasoning, and mathematical reasoning.
- Stable training: Based on curriculum-guided reinforcement learning and difficulty-aware retrospective sampling, ensuring a stable training process.
- Hybrid reward: Combines rule-based and model-based rewards to balance precision and recall.
- Broad applicability: Suitable for a variety of practical application scenarios, such as legal document analysis, financial report interpretation, and scientific paper reading.
- High performance: On multiple long-text document question answering (DocQA) benchmarks, it outperforms existing flagship models such as OpenAI-o3-mini and Qwen3-235B-A22B.
Technical principles of QwenLong-L1-32B
- Progressive context extension: Divides the training process into multiple stages, gradually increasing the context length to ensure that the model can stably adapt to longer contexts at each stage. Sampling is performed according to sample difficulty, prioritizing complex samples to encourage the model to conduct deeper exploration.
- Hybrid reward mechanism: Based on strict matching of the final answer and format validation, ensuring the precision of the model output. A small language model is used as an evaluator to judge the semantic equivalence between the generated answer and the standard answer, improving the model's recall.
- Reinforcement learning algorithm: Optimizes the policy based on group relative advantage estimation, avoiding the use of an additional value network and reducing computational complexity. It combines high clipping thresholds, dynamic sampling strategies, token-by-token loss, and overlong reward shaping to ensure a more stable and efficient RL process.
- Pretraining and fine-tuning: Uses pretrained short-text reasoning models (such as R1-Distill-Qwen-14B and R1-Distill-Qwen-32B) as base models. Before RL training, supervised fine-tuning is performed based on high-quality annotated data to provide the model with a robust initial policy.
Project addresses of QwenLong-L1-32B
- GitHub repository：https://github.com/Tongyi-Zhiwen/QwenLong-L1
- HuggingFace model library：https://huggingface.co/Tongyi-Zhiwen/QwenLong-L1-32B
- arXiv technical paper：https://arxiv.org/pdf/2505.17667
Application scenarios of QwenLong-L1-32B
- Legal field: Analyzes legal documents, extracts key information, answers complex legal questions, and supports legal case analysis and judgment prediction.
- Financial field: Processes financial reports, performs data analysis and forecasting, and supports financial decision-making and risk management.
- Scientific research field: Extracts experimental results and conclusions from scientific research papers, assisting scientific research and academic writing.
- Education field: Aids teaching, provides personalized learning content and answers, supports online courses and intelligent tutoring.
- Intelligent Customer Service: Handles complex user inquiries, provides accurate answers and suggestions, supports customer service in fields such as finance and technical support.