DeepSeek: DeepSeek V3.2

DeepSeek V3.2 is DeepSeek’s text model for reasoning and agent work with an emphasis on efficient long-context processing. The producer organises its main changes around DeepSeek Sparse Attention, a scaled reinforcement-learning framework and a pipeline for training agentic tasks. It accepts written questions and context and produces text, including an answer, code or tool-call content. It is distinct from the earlier V3.2 Exp release and the separate Speciale reasoning variant.

DeepSeek Sparse Attention adds an indexer and fine-grained token selection to the latent-attention architecture. For a long sequence, the indexer selects the earlier token information used by the main attention calculation. The report says this is the architectural change from V3.1 Terminus and uses the same architecture as V3.2 Exp. Its published context is 128K tokens, which supports substantial source material and a continuing task, while generated reasoning and tool history also consume the available context.

Post-training covers both thinking and non-thinking responses across mathematics, programming, logical reasoning, general agent tasks, agentic coding and search. Domain specialists supply data that is consolidated into the final model, followed by mixed reinforcement learning. For an ordinary question, a direct answer may be appropriate; for a difficult problem, the thinking workflow can develop a solution before presenting the final result. Specify the required answer or artifact independently of the intermediate reasoning.

A substantial edition change is thinking with tools. The updated chat format allows reasoning to continue through a tool-calling task rather than treating analysis and action as unrelated stages. A model-selected operation is executed by the application, which returns its result for the next step. The producer’s tool-context rule retains historical reasoning while only tool-related messages are added; when a new user message arrives, it discards that reasoning while preserving tool calls and results. Correctly distinguished conversation roles matter to that behaviour.

For coding, provide the source fragment or relevant files, the observed problem, the expected result and the conditions the change must preserve. An illustrative brief is: “Use the supplied code and tool results to locate the cause, propose the smallest correction and return the final patch with its supporting validation result.” For a search task, describe the question, required evidence and what counts as a completed answer. Search and execution tools are supplied by the surrounding application rather than by the text alone.

The producer recommends temperature 1.0 and top-p 0.95 for local use. V3.2 introduces a revised tool-call representation and a dedicated message encoder. Its developer role is reserved for the producer’s search-agent template, and the official API does not accept that role. The reference parser expects well-formed output and does not repair malformed results; when an application uses generated fields or tool arguments, it should check their structure before acting on them.

The technical report identifies long reasoning trajectories and redundant self-verification as practical limitations. An agent task can reach the 128K boundary even when the original source material fits, because tool outputs and repeated reasoning accumulate. Keep a concise record of relevant observations and make the final artifact explicit. For knowledge-sensitive work, attach the authoritative material or provide a suitable search tool rather than expecting training knowledge to define every fact.

DeepSeek V3.2 Speciale is a separate high-compute deep-reasoning variant and does not support tool calling in the model card. Its features should not be used as the standard V3.2 tool profile. Likewise, the later V4 generation and the experimental predecessor have separate identities and configuration choices. Retain the named V3.2 edition when comparing tasks whose model and prompting format matter.

For a practical source-based analysis, label the materials and ask for a conclusion supported by the relevant passages, with assumptions and missing information kept distinct. For an agent, state the allowed functions, the expected final result and the completion criterion. This prepares the model’s reasoning-and-tool workflow around a reviewable text, code or analytical artifact tied to the supplied evidence.