← All guidance

Prompting DeepSeek V3.2

How to prompt DeepSeek V3.2, an open-weight reasoning-and-general model with a 1M-token context. Read the shared Prompting guides principles first; this guide carries only what is distinctive to V3.2. This is reference for when you target, self-host, or evaluate DeepSeek. (DeepSeek V4, released April 2026, is the newer flagship; V3.2 remains a widely-deployed open-weight baseline.)

When to reach for it

DeepSeek V3.2 is a strong open-weight pick for multi-step reasoning and maths, and a common self-hosted baseline where its permissive weights and low inference cost matter. Reach for it as an open reasoning baseline or when comparing reasoning traces across models.

Prompting principles

• Don't add chain-of-thought scaffolding. It is a reasoning model — it already deliberates internally. "Think step by step" instructions add verbosity and can distract it; DeepSeek's own guidance discourages them on the reasoning path.
• Set temperature by task. DeepSeek publishes task-specific temperatures: 0.0 for coding and maths (determinism), 1.0 for general conversation, higher (1.3) for translation, and higher still for creative writing. The API also remaps temperature internally, so tune against observed behaviour.
• Constrain JSON tightly. For structured output, instruct JSON-only in the system prompt, give a tight schema, avoid stray delimiter sequences, and lower the temperature.
• Put documents before the question. For search/RAG over supplied documents, place the document block first and the instruction after it.