DeepSeek-V3.2 handles long-context inputs by replacing the dense self-attention in V3.1 with DeepSeek Sparse Attention (DSA), which is a two-stage sparse attention process: a lightweight “indexer” that picks promising tokens, and a fine-grained selector that attends more densely within that subset. This trims the quadratic scaling of classic attention down to something closer to linear in practice, especially when many tokens are only weakly relevant. In vLLM’s reference configuration, the V3.2-Exp checkpoint is configured for 128k context length, and the combination of DSA plus vLLM’s paged attention gives a big cut in memory and compute compared with running the same context through V3.1.
During inference, you still pay two main costs: prefill (processing the whole prompt) and decode (generating new tokens). DSA primarily improves the prefill side for large prompts by avoiding full attention across all tokens, while trying to keep behavior close enough to dense attention that benchmarks and outputs remain similar. DeepSeek’s announcement and independent write-ups report that long-context API calls can see up to roughly half the cost compared with V3.1 at similar quality, which lines up with what you’d expect from sparse attention plus implementation work in vLLM and SGLang. The important caveat is that “handles” doesn’t mean “magically reasons over 100k+ tokens perfectly”—you still see performance decay on very old context, and very noisy prompts can confuse the model even if they fit.
Because of those practical limits, you usually get better results by combining V3.2 with retrieval rather than just dumping everything into context. A natural pattern is to store your corpus in a vector database like Milvus or Zilliz Cloud, and then use long context only for what truly benefits from it: keeping a decent chat history, carrying multi-step tool traces, or stitching together a handful of long documents when you really must. For example, you might keep 4–8k tokens of chat history plus 10–20 retrieved passages from Milvus in the prompt, instead of streaming in hundreds of thousands of raw tokens. This way, you leverage DSA’s efficiency when needed but mostly rely on fast, targeted vector search to keep prompts small, relevant, and cheap.