DeepSeek: DeepSeek V3.2 Exp

DeepSeek V3.2 Exp is DeepSeek’s experimental text-model edition built on V3.1 Terminus. Its defining change is DeepSeek Sparse Attention, introduced to study more efficient training and inference for extended text sequences. The task remains text-based: written instructions and context produce a text answer, code or agent-related output. This is the sparse-attention research edition, separate from the final V3.2 model and its later thinking-with-tools format.

DSA uses a lightweight indexer and fine-grained token selection so the main attention calculation focuses on selected earlier token information. DeepSeek deliberately aligned the experimental training configuration with V3.1 Terminus to isolate the architectural change. The V3.2 report confirms that the final V3.2 and V3.2 Exp share this architecture and describes continued training from the 128K-context Terminus base. The published experimental profile uses that 128K long-context setting.

The producer’s experimental card keeps reasoning and agentic work in scope while preserving the prior model’s task profile. Its evidence covers reasoning without tools and agentic tool use as distinct workflows. Prepare a clear question, relevant source context and the desired final form. If the task uses a tool, the application supplies its definition and returns the result; the model’s proposed action is not the same as an operation that has already executed.

The final V3.2 card introduces a revised tool-call format and explicitly describes thinking with tools. Those updates belong to the final release and should not be treated as the experimental edition’s entire interaction contract. For V3.2 Exp, retain the edition’s own message formatting and the inherited Terminus task context rather than assuming that a similar name makes every later control or role interchangeable.

For a long document comparison, label each source and identify the sections relevant to the question. An illustrative brief is: “Compare these requirements with the proposed implementation, identify each mismatch and cite the passage that establishes it. End with a prioritised correction list.” For code, provide the failing behaviour, relevant fragment and expected result, then ask for a focused change with a clear explanation. The source layout and completion rule help direct a long-context task toward a usable artifact.

The producer’s reference implementation includes the sparse-attention indexer, and its card records an important implementation correction: indexer RoPE uses a non-interleaved layout, whereas the latent-attention module expects an interleaved layout. This is a producer-reference detail relevant to faithful model execution, rather than a different user prompt or a new capability. It also means a model name alone does not describe every implementation’s handling of the experimental architecture.

A continuing agent task must budget its source text, conversation, generated reasoning and tool observations together. Keep the useful observations organised and ask for the final answer to distinguish evidence, assumptions and proposed actions. For a source-dependent factual result, supply the authoritative material or a suitable tool rather than relying on broad training knowledge. The output remains a text, code or analytical record that can be checked against the task and its inputs.