18:I[66739,["/_next/static/chunks/0u0nyy7wb9me2.js?dpl=dpl_EVxQmm5ee6SfSwpTLEi7dkcqj48i","/_next/static/chunks/05hx6kuf6-oy4.js?dpl=dpl_EVxQmm5ee6SfSwpTLEi7dkcqj48i","/_next/static/chunks/12hmyxte1w1dw.js?dpl=dpl_EVxQmm5ee6SfSwpTLEi7dkcqj48i","/_next/static/chunks/166.u_60qj09..js?dpl=dpl_EVxQmm5ee6SfSwpTLEi7dkcqj48i","/_next/static/chunks/0g9_wl34k-6l4.js?dpl=dpl_EVxQmm5ee6SfSwpTLEi7dkcqj48i","/_next/static/chunks/06dn~uer6.43l.js?dpl=dpl_EVxQmm5ee6SfSwpTLEi7dkcqj48i"],"default"] 1a:I[38632,["/_next/static/chunks/0qqtp3r0h7kah.js?dpl=dpl_EVxQmm5ee6SfSwpTLEi7dkcqj48i","/_next/static/chunks/0iukvycbd42rl.js?dpl=dpl_EVxQmm5ee6SfSwpTLEi7dkcqj48i"],"IconMark"] 19:T52e,

The paper introduces the DeepSeek-V3.2 Large Language Model (LLM) framework, explicitly designed to bridge the performance gap between open-source and proprietary systems. A key technical advancement is the DeepSeek Sparse Attention (DSA) mechanism, which significantly improves efficiency by reducing computational complexity for processing long-context sequences. The model's reasoning and agentic proficiencies were enhanced through a scalable reinforcement learning framework that allocates substantial post-training compute and a novel synthesis pipeline for generating large-scale agentic tasks. DeepSeek-V3.2 achieves performance parity with closed-source models like GPT-5 on standard benchmarks, demonstrating strong tool-use and generalization capabilities. Notably, the high-compute variant, DeepSeek-V3.2-Speciale, achieved gold-medal performance in the 2025 International Mathematical Olympiad (IMO) and other top-tier competitions, reaching capability parity with models such as Gemini-3.0-Pro. Overall, the work establishes a new performance milestone for open LLMs, though the authors note challenges in world knowledge and token efficiency remain.

DeepSeek-V3.2: Gold-Medal AI via Sparse Attention

Steven AI Talk

DeepSeek-V3.2: Gold-Medal AI via Sparse Attention