DeepSeek V3.2 Exp

DeepSeek V3.2 Exp is an experimental release positioned as the immediate successor to V3.1-Terminus, carrying forward that model's lineage while introducing a new attention mechanism. The defining change is DeepSeek Sparse Attention (DSA), a fine-grained sparse attention scheme designed to accelerate both training and inference over long contexts while keeping output quality close to the dense baseline. Because DSA preserves generation behavior on standard tasks, the release functions as a drop-in experimental upgrade for teams already running V3.1-Terminus in production, with the prior checkpoint kept on a temporary comparison endpoint through mid-October 2025 to ease A/B testing.

In practice, V3.2 Exp is aimed at workloads that benefit from long-context throughput, such as multi-document analysis, codebases spanning many files, and extended agent loops, where sparse attention reduces compute cost per token without a noticeable quality drop on reported benchmarks. DeepSeek positions the new model as benchmark-parity with V3.1-Terminus, framing DSA primarily as an efficiency story rather than a capability leap, and pairs the release with a 50%+ reduction in API pricing to encourage migration. For practitioners, it is a useful choice when the goal is to retain familiar DeepSeek response quality at a lower cost and with better long-context scaling, while teams that need maximum reasoning depth may still prefer specialized reasoning variants in the broader family.