推理模型在写出最终答案之前,先在中间推理符号中处理问题。这改善了数学、代码和规划等多步骤任务的结果,使用更多输出标记和时间,推理努力等设置对其进行控制。
标准语言模型会立即开始写出答案。推理模型首先生成一系列中间步骤,通常称为思考,然后写出答案。在这个阶段,它可以将问题拆分成部分,进行计算,测试一种方法,并在最终做出结果之前自我修正。
推理是一种能力,而不是独立的模型类型。文本模型也可以是推理模型,很多模型允许你为每个请求开启或关闭推理。有些模型总是推理,通常可以调整推理的程度。
推理能够改善包含多个相互依赖步骤的任务的结果:
它对简单的查找、分类、短暂重写或日常聊天几乎没什么帮助,主要增加延迟和输出令牌。对于这些任务,可以降低工作量或关闭推理。
推理模型的提示最好用的是,当它们明确目标、限制条件和预期答案格式,并将规划工作交给模型时。详细的step-by-step说明通常没必要,因为推理阶段的工作就是按步骤来做的。
EmpirioLabs AI 显示了三个控件,每个型号都接受本页列出的控件:
推理(e) 设置深度,使用命名级别,例如: 低, (中文), 媒介 以及 高.在许多型号上, 无 关闭推理功能。可用的级别因型号而异。thinking_budget 设置模型在思考开启时可能花费在推理上的最多代币。如果你省略它,许多模型会自行决定推理的时间。启用 思考 在提供选择的型号上,可以开关推理。这些参数在 /v1/chat/completions, (中文), /v1/responses (英语) 以及 /v1/messages (英语) 适用于支持每个端点的模型。开启 /v1/messages (英语) 你也可以发送 思考 对象: 已启用 与 budget_tokens 设定预算, 自适应 无需佩戴就能启动思考,并且 已禁用 将其关闭。在某些型号上,有一个请求使用 response_format 或者强制工具调用响应,除非它设置了 启用 思考.参见 跨端点的模型参数 详情。
推理是生成的文本,因此即使你从未看到它也会计费。在大多数模型中,推理令牌作为输出令牌计费,每个模型页面都会说明其计费方式。有些模型返回推理文本,例如在 推理(c) 在聊天完成中,而其他则仅返回答案。在 /v1/responses (英语), (中文), usage.output_tokens_details.reasoning_tokens 报告使用了多少辆。在某些型号上, preserve_thinking 将先前助手的推理发送回模型,转回模型 /v1/chat/completions,以及早期推理将作为输入代币来计费。
推理也计入产出极限。 max_completion_tokens 大写推理和答案结合使用,在某些模型上也是如此 max_tokens 同样如此。当限制太低时,模型可以用来推理,结果给出简短或空洞的答案,所以两者都留有空间。
更高的努力也意味着在回答开始前等待时间更长。流式播放回复以显示进展,对于超过15分钟的回复,请使用 WebSocket 模式 /v1/responses (英语).比较具备推理能力的 文本模型,或者尝试不同的努力强度 Playground。 。 。.
共 65 个模型。每个模型的页面都有价目表、参数和代码。
| 模型 | 参数 | 范围 / 值 |
|---|---|---|
| DeepSeek V4 Flash 0731 DeepSeek | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max; 默认 medium thinking_budget: 1 到 393216; 默认 32768 enable_thinking: 默认 true |
| DeepSeek V4 Pro DeepSeek | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max; 默认 medium thinking_budget: 1 到 393216; 默认 32768 enable_thinking: 默认 true |
| DeepSeek V4 Pro 0813 DeepSeek | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max; 默认 medium thinking_budget: 1 到 393216; 默认 32768 enable_thinking: 默认 true |
| DeepSeek V4.1 Flash 1017 DeepSeek | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max; 默认 medium thinking_budget: 1 到 393216; 默认 32768 enable_thinking: 默认 true |
| GLM 5.1 Z.ai | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max; 默认 medium thinking_budget: 1 到 38912; 默认 32768 enable_thinking: 默认 true |
| Gemma 4 26B-A4B Google | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max; 默认 medium thinking_budget: 128 到 32768; 默认 4096 enable_thinking: 默认 true |
| Kimi K2.6 Moonshot AI | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max; 默认 medium thinking_budget: 1 到 81920; 默认 32768 enable_thinking: 默认 true |
| Qwen3.5 122B-A10B Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max thinking_budget: 1 到 81920 enable_thinking: 默认 true |
| 模型 | 参数 | 范围 / 值 |
|---|---|---|
| Qwen3.5 27B Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max thinking_budget: 1 到 81920 enable_thinking: 默认 true |
| Qwen3.5 35B-A3B Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max thinking_budget: 1 到 81920 enable_thinking: 默认 true |
| Qwen3.5 397B-A17B Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max thinking_budget: 1 到 81920 enable_thinking: 默认 true |
| Qwen3.5 4B Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max; 默认 medium thinking_budget: 1024 到 32768; 默认 4096 enable_thinking: 默认 true |
| Qwen3.5 9B Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max; 默认 medium thinking_budget: 1024 到 32768; 默认 4096 enable_thinking: 默认 true |
| Qwen3.6 27B Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max thinking_budget: 1 到 131072 enable_thinking: 默认 true |
| Qwen3.6 Flash Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max; 默认 medium thinking_budget: 1 到 64000; 默认 32768 enable_thinking: 默认 true |
| Qwen3.6 Plus Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max thinking_budget: 1 到 131072 enable_thinking: 默认 true |
| Qwen3.7 Flash Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max thinking_budget: 1 到 131072 enable_thinking: 默认 true |
| Qwen3.7 Max Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max thinking_budget: 1 到 262144 enable_thinking: 默认 true |
| Qwen3.7 Plus Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max thinking_budget: 1 到 262144 enable_thinking: 默认 true |
| Qwen3.8 Flash Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max thinking_budget: 1 到 262144 enable_thinking: 默认 true |
| Qwen3.8 Max Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max thinking_budget: 1 到 262144 enable_thinking: 默认 true |
| Qwen3.8 Max 0902 Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max thinking_budget: 1 到 262144 enable_thinking: 默认 true |
| Qwen3.8 Omni Flash Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: none, low, medium, high, max thinking_budget: 1 到 262144 enable_thinking: 默认 true |
| Qwen4 Max Alibaba Cloud | reasoning_effort thinking_budget enable_thinking | reasoning_effort: low, high, xhigh; 默认 high thinking_budget: 1 到 262144 enable_thinking: 默认 true |
| GLM 5.2 Z.ai | reasoning_effort enable_thinking | reasoning_effort: none, minimal, low, medium, high, xhigh, max; 默认 max enable_thinking: 默认 true |
| Kimi K3.1 Pro Moonshot AI | reasoning_effort enable_thinking | reasoning_effort: low, medium, high, max; 默认 max enable_thinking: 默认 true |
| Nova Lite 2 Amazon | reasoning_effort enable_thinking | reasoning_effort: low, medium, high; 默认 medium enable_thinking: 默认 true |
| Qwen3.5 Flash Alibaba Cloud | thinking_budget enable_thinking | thinking_budget: 1 到 81920 enable_thinking: 默认 true |
| Qwen3.5 Plus Alibaba Cloud | thinking_budget enable_thinking | thinking_budget: 1 到 81920 enable_thinking: 默认 true |
| Qwen3.6 35B A3B Alibaba Cloud | reasoning_effort enable_thinking | reasoning_effort: none, low, medium, high, max; 默认 medium enable_thinking: 默认 true |
| Qwen3.8 27B Alibaba Cloud | reasoning_effort enable_thinking | reasoning_effort: none, low, medium, xhigh; 默认 xhigh enable_thinking: 默认 true |
| Seed 2.0 Code ByteDance | reasoning_effort enable_thinking | reasoning_effort: low, medium, high; 默认 medium enable_thinking: 默认 true |
| Seed 2.0 Lite ByteDance | reasoning_effort enable_thinking | reasoning_effort: low, medium, high; 默认 medium enable_thinking: 默认 true |
| Seed 2.0 Mini ByteDance | reasoning_effort enable_thinking | reasoning_effort: low, medium, high; 默认 medium enable_thinking: 默认 true |
| Seed 2.0 Pro ByteDance | reasoning_effort enable_thinking | reasoning_effort: low, medium, high; 默认 medium enable_thinking: 默认 true |
| Seed 2.1 Turbo ByteDance | reasoning_effort enable_thinking | reasoning_effort: low, medium, high; 默认 high enable_thinking: 默认 true |
| Fugu Max Sakana AI | reasoning_effort | high, xhigh, max; 默认 high |
| Fugu Ultra v1.0 Sakana AI | reasoning_effort | high, xhigh, max; 默认 high |
| Fugu Ultra v1.1 Sakana AI | reasoning_effort | high, xhigh, max; 默认 high |
| Fugu Ultra v2.0 Sakana AI | reasoning_effort | high, xhigh, max; 默认 high |
| GLM 4.5 Flash Z.ai | enable_thinking | 默认 true |
| GLM 4.6V Flash Z.ai | enable_thinking | 默认 true |
| GLM 4.6V FlashX Z.ai | enable_thinking | 默认 true |
| GLM 4.7 Flash Z.ai | enable_thinking | 默认 true |
| GLM 4.7 FlashX Z.ai | enable_thinking | 默认 true |
| GLM 5.3 Z.ai | reasoning_effort | low, high, max; 默认 max |
| GLM 5.3 Flash Z.ai | reasoning_effort | low, high, max; 默认 max |
| Gemini 3.8 Live Extended Thinking Google | reasoning_effort | low, medium, high; 默认 medium |
| Kimi K3 Moonshot AI | reasoning_effort | low, high, max; 默认 max |
| MiMo V2.6 Flash Xiaomi | enable_thinking | 默认 true |
| MiMo V2.6 Pro Xiaomi | enable_thinking | 默认 true |
| MiMo V2.6 Pro UltraSpeed Xiaomi | enable_thinking | 默认 true |
| MiniMax M3 MiniMax | enable_thinking | 默认 true |
| MiniMax M3.1 Flash MiniMax | reasoning_effort | low, medium, high, xhigh, max; 默认 max |
| Muse Glimmer 30B Meta AI | reasoning_effort | none, low, medium, high, max; 默认 high |
| Muse Spark 1.1 Meta AI | reasoning_effort | minimal, low, medium, high, xhigh; 默认 medium |
| Muse Spark 1.2 Meta AI | reasoning_effort | minimal, low, medium, high, xhigh; 默认 medium |
| Muse Spark 1.3 Meta AI | reasoning_effort | minimal, low, medium, high, xhigh; 默认 medium |
| Perplexity Advanced Deep Research Perplexity | reasoning_effort | low, medium, high; 默认 medium |
| Perplexity Deep Research Perplexity | reasoning_effort | low, medium, high; 默认 medium |
| Perplexity Pro Search Perplexity | reasoning_effort | minimal, low, medium, high; 默认 minimal |
| Perplexity Sonar Reasoning Pro Perplexity | reasoning_effort | low, medium, high; 默认 medium |
| Step 3.5 Flash 2603 StepFun | reasoning_effort | low, high; 默认 low |
| Step 3.7 Flash StepFun | reasoning_effort | low, medium, high; 默认 low |
| Step 5 Preview StepFun | reasoning_effort | low, medium, high; 默认 low |
是的。推理是生成的文本,所以即使模型没有返回推理,它也会被计费。在大多数模型中,它作为输出代币计费,每个模型页面都会说明具体计费方式。
在许多模型上,是的:根据模型提供的控制,将enable_thinking设置为假或reasoning_effort为无。有些模型总是会推理,只允许你更改努力程度。
reasoning_effort选择一个命名的深度层级,比如低或高。thinking_budget设定在思考启动时用于推理的最大标记数。
当输出极限过低时,模型可以把全部精力都用在推理上。提高max_completion_tokens或max_tokens,或降低推理努力,这样就为答案留出空间。