Deepseek V3.2 Exp

DeepSeek-V3.2-Exp, released September 2025, introduced DeepSeek Sparse Attention (DSA), a fine-grained token-selection mechanism that lowers attention cost from O(L squared) toward O(Lk) for long sequences. The 671-billion-parameter model keeps 37 billion active parameters and supports a 163,840-token context for large-document processing. A lightweight indexer scores context tokens, after which sparse attention runs over only the top-k most relevant tokens, sharply reducing long-context compute while keeping benchmark quality on par with V3.1-Terminus. It offers both an explicit chain-of-thought thinking mode and a faster non-thinking mode. Released open-weight under an MIT license, V3.2-Exp serves as a research and deployment testbed for cost-effective long-context reasoning, suited to document-heavy applications and agent systems handling variable-length inputs.