The main architecture change that differentiates DeepSeek-V3.2-Exp from DeepSeek-V3.1-Terminus is the introduction of DeepSeek Sparse Attention (DSA) on top of an otherwise similar Mixture-of-Experts backbone. The model card is explicit that training configurations were deliberately aligned between V3.2-Exp and V3.1-Terminus so that any differences in behavior can be attributed to DSA rather than to extra parameters or radically different hyperparameters. Conceptually, you can think of V3.2-Exp as “V3.1 with smarter attention”: it still has a large MoE transformer with a 128K context window, but the way it decides which past tokens to focus on is more efficient and more selective, especially for long sequences.
DSA itself has two key components: a “lightning indexer” and a fine-grained token selection module. The lightning indexer computes relevance scores between each query token and previous tokens using a lightweight attention-like mechanism that can run in low precision (FP8) for speed. Then, for each query token, the model keeps only the top-k key/value pairs according to those scores and runs standard attention on that subset instead of on the full sequence. This changes the dominant cost from O(L²) dense attention to something closer to O(L·k), with k much smaller than L, while the indexer itself is designed to be cheaper than the original dense block. Training is done in two phases: first, a dense warm-up where the indexer is trained to match dense attention distributions; then, continued pre-training with sparse attention enabled.
For developers, the important point is that V3.2-Exp behaves like V3.1 in terms of capabilities but scales better with sequence length, which matters for workloads like log analysis, long-document QA, and complex RAG. You can send up to 128K tokens to deepseek-chat or deepseek-reasoner through the API, but thanks to DSA you pay less in latency and cost when you approach those limits. In practice, you’ll still want to front-load relevance using a vector database such as Milvus or Zilliz Cloud so you aren’t wasting DSA’s budget on irrelevant tokens. That pattern—vector search for recall, DSA-powered transformer for reasoning—is exactly what you see in public tutorials that combine DeepSeek models with Milvus-based RAG pipelines.