Source: DeepSeek API Docs — Thinking Mode

DeepSeek's reference for thinking mode: enabling it, effort mapping, reasoning_content, multi-turn and tool-call rules, ignored sampling params.

Contents
  • What: official API documentation page
  • Publisher: DeepSeek
  • Date: accessed 2026-11-14 (models deepseek-flash, deepseek-v4-pro)
  • Original: Thinking mode

Summary

How to use DeepSeek’s reasoning through OpenAI-, Anthropic- and Responses-compatible endpoints. Unlike the closed vendors, the reasoning text itself is returned.

Key claims

  • Thinking is on by default with effort high. OpenAI format: thinking: {type: enabled|disabled} plus reasoning_effort low|high|max. Other effort names are mapped onto these three.
  • The chain of thought is returned in reasoning_content, next to content (delta.reasoning_content when streaming).
  • Without tools, earlier turns’ reasoning_content is ignored if sent. With tools, it must be passed back for all previous turns, or the API returns a 400.
  • In thinking mode the model can run several reasoning/tool-call rounds before answering.
  • temperature, presence_penalty and frequency_penalty are silently ignored in thinking mode; top_p is clamped or fixed.

Feeds

Reasoning models · Extended thinking

Suggest changes →