Unverified60% confidenceFactExact time
Reasoning Effort（推理努力度）是DeepSeek R1系列模型引入的重要参数，用于控制模型在生成最终答案前执行的思维链推理深度和广度
2
Sources
60%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026
Sources
DeepSeek-Reasonix: An AI Coding Agent Built for DeepSeek That Cuts Costs by 80%
bilibili程序员晓刘5/27/2026
Related Claims
VerifiedReasoning Effort本质上是控制模型思维链深度的参数，档位越高模型允许的内部推理步骤越多，延迟和成本也成比例上升78% similarVerified推理模型以OpenAI o1、DeepSeek-R1、QwQ为代表，核心机制是思维链扩展，在给出最终答案前生成大量中间推理步骤78% similarUnverifiedo1、DeepSeek-R1等模型将思维链内化为训练目标，在输出答案前生成大量内部草稿推理77% similarUnverified深度思考模式下，模型会先在内部生成中间推理步骤，再基于这些步骤得出最终结论，在处理多步逻辑推导任务时表现显著优于直接生成75% similarUnverified推理模型（如OpenAI的o1/o3系列、Anthropic的Claude with extended thinking、DeepSeek-R1）在生成最终答案前会先输出思考过程（thinking tokens）75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/46112API
curl https://kongchang.com/api/v1/knowledge/claims/46112MCP
get_claim(id=46112)