DeepSeek V3.2 is a large language model and AI assistant optimized for both everyday chat and advanced multi-step reasoning. It introduces a dual-mode workflow, allowing users to choose between a fast conversational mode and a slower, deliberate thinking mode that integrates deeply with tool use for complex tasks such as math, code generation, and structured planning.[1][2] Built on the DeepSeek Sparse Attention mechanism, it delivers efficient long-context handling with reduced compute cost while maintaining high output quality.[1][3][5] The V3.2 family also includes the V3.2-Speciale research variant, which reaches gold-medal performance on competitions like IMO and IOI and targets heavy reasoning and agent workloads.[1][2][3]