A.I.
DeepSeek · current
DeepSeek thinking mode
Before the final answer, DeepSeek can emit chain-of-thought reasoning. Thinking mode is on by default (default effort high). You can turn it off or change effort.
- Status
- Feature, a setting on Flash and V4 Pro
Controls
| Goal | OpenAI-style Chat Completions | Notes |
|---|---|---|
| Enable / disable | thinking.type enabled or disabled | Pass via extra_body in the OpenAI SDK |
| Effort | reasoning_effort: low, high, or max | Docs map effort for Flash and Pro the same |
| Responses API | reasoning.effort: none / low / high / max | none disables thinking |
Limits
Thinking mode does not support temperature, top_p, presence_penalty, or frequency_penalty (ignored for compatibility). Tool calls work with thinking; you must pass reasoning_content back on later turns or the API returns 400.