DeepSeek-AI, A. Liu, Aoxue Mei, B. Lin, Bing Xue, Bing-Li Wang, Bin Xu, Bochao Wu et al.
DeepSeek-V3.2 is an open LLM that enhances both efficiency and reasoning/agent performance through sparse attention, reinforcement learning scaling, and an agentic task synthesis pipeline.
Existing open LLMs suffer from high computational costs in long contexts and limited performance in complex reasoning and tool-use scenarios.
(1) DeepSeek Sparse Attention (DSA) reduces computational complexity while preserving performance in long contexts. (2) Scaling post-training compute via a robust reinforcement learning framework achieves reasoning capability comparable to GPT-5. (3) A large-scale agentic task synthesis pipeline generates training data for tool-use scenarios, improving generalization and instruction-following robustness through agentic post-training.
DeepSeek-V3.2 performs comparably to GPT-5, and its high-compute variant (Speciale) surpasses GPT-5 while matching Gemini-3.0-Pro in reasoning, achieving gold-medal performance at the 2025 IMO and IOI.