DeepSeek-V3.2 is a leading model that excels in reasoning, complex reasoning capabilities, and efficient agent scenarios. Built on the DeepSeek Sparse Attention (DSA) mechanism, it significantly reduces computational overhead while optimizing long-context performance. Through a scalable reinforcement learning framework, its overall capabilities reach the level of GPT-5, with the high-compute variant V3.2-Speciale approaching Gemini-3.0-Pro in reasoning performance. Additionally, leveraging a large-scale synthetic agent task tower, the model demonstrates stronger tool-calling and multi-step decision-making abilities, achieving gold-medal-level results in both the 2025 IMO and IOI.