We introduce DeepSeek-V3.2, a model that harmonizes high computational efficiency with superior reasoning and agent performance. The key technical breakthroughs of DeepSeek-V3.2 are as follows: (1) DeepSeek Sparse Attention (DSA): We introduce DSA, an efficient attention mechanism that substantially reduces computational complexity while preserving model performance in long-context scenarios. (2) Scalable Reinforcement Learning Framework: By implementing a robust reinforcement learning protocol and scaling post-training compute, DeepSeek-V3.2 performs comparably to GPT-5. Notably, our high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and exhibits reasoning proficiency on par with Gemini-3.0-Pro, achieving gold-medal performance in both the 2025 International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI). (3) Large-Scale Agentic Task Synthesis Pipeline: To integrate reasoning into tool-use scenarios, we developed a novel synthesis pipeline that systematically generates training data at scale. This methodology facilitates scalable agentic post-training, yielding substantial improvements in generalization and instruction-following robustness within complex, interactive environments.
核心贡献 · Key contributions
提出 DeepSeek 稀疏注意力（DSA），将注意力复杂度从 O(L^2)降至 O(Lk)，同时保持长上下文性能。 Introduces DeepSeek Sparse Attention (DSA), reducing attention complexity from O(L^2) to O(Lk) while preserving long-context performance.
开发可扩展强化学习框架，后训练算力超过预训练的 10%，实现 GPT-5 级别的推理能力。 Develops a scalable reinforcement learning framework with over 10% of pre-training compute for post-training, achieving GPT-5-level reasoning.
DeepSeek-V3.2-Speciale 超越 GPT-5，媲美 Gemini-3.0-Pro，在 2025 年 IMO 和 IOI 中获得金牌。 DeepSeek-V3.2-Speciale surpasses GPT-5 and matches Gemini-3.0-Pro, winning gold medals in IMO 2025 and IOI 2025.
提出离策略序列掩码和保持路由策略，稳定 MoE 模型的强化学习训练。 Proposes off-policy sequence masking and keep routing to stabilize RL training for MoE models.
通过冷启动和上下文管理将推理融入工具使用，提升智能体指令遵循能力。 Integrates reasoning into tool-use via cold-start and context management, improving agentic instruction following.
局限 · Limitations
由于总训练算力较少，世界知识广度落后于领先的专有模型。 World knowledge breadth lags behind leading proprietary models due to fewer total training FLOPs.
词元效率较低；DeepSeek-V3.2 需要更长的生成轨迹才能匹配 Gemini-3.0-Pro 的输出质量。 Token efficiency is lower; DeepSeek-V3.2 requires longer generation trajectories to match output quality of Gemini-3.0-Pro.
复杂任务解决能力仍不如前沿模型，需进一步改进基础模型和后训练。 Complex task solving still inferior to frontier models, motivating further refinement of foundation model and post-training.
智能体工作流常触及 128K 上下文限制，需要上下文管理策略才能充分扩展测试时算力。 Agentic workflows often hit 128K context limit, requiring context management strategies for full test-time compute scaling.
非思考模式性能略逊于思考模式，但仍具竞争力。 Non-thinking mode performance is slightly worse than thinking mode, though still competitive.
论文章节 · Sections（共 17）
摘要Abstract
1 引言1 Introduction
2.1 DeepSeek 稀疏注意力2.1 DeepSeek Sparse Attention
2.2 奇偶性评估2.2 Parity Evaluation
2.3 推理成本2.3 Inference Costs
3 后训练3 Post-Training
3.1 扩展 GRPO3.1 Scaling GRPO
3.2 工具使用中的思考3.2 Thinking in Tool-Use
4.1 主要结果4.1 Main Results
4.2 DeepSeek-V3.2-Speciale 结果4.2 Results of DeepSeek-V3.2-Speciale
4.3 综合智能体任务4.3 Synthesis Agentic Tasks
4.4 搜索智能体的上下文管理4.4 Context Management of Search Agent
5 结论、局限与未来工作5 Conclusion, Limitation, and Future Work
附录 A MLA 的 MHA 和 MQA 模式Appendix A MHA and MQA Modes of MLA
附录 C 非思考型 DeepSeek-V3.2 智能体评估Appendix C Non-thinking DeepSeek-V3.2 Agentic Evaluation
附录 D IOI、ICPC 世界总决赛、IMO 和 CMO 的评估方法Appendix D Evaluation Method of IOI, ICPC World Final, IMO, and CMO