DeepSeek-V3.2

在 DeepSeek-V3.2 上进行对话补全，OpenAI 兼容端点：deepseek-reasoner 连同推理过程一起返回答案，deepseek-chat 直接回答。

POST/v3/llm/deepseek-3p2/chat/completions试一试

• 价格0.01 / 0.02 IC每 1k 词元，输入 / 输出
• 累计调用28.4K次

• v3.2
• 运行中

输入

model

deepseek-reasoner

messages

role: user content: สวัสดีครับ คุณช่วยอะไรได้บ้าง?

max_tokens

4096

temperature

0.7

top_p

0.9

输出POST /v3/llm/deepseek-3p2/chat/completions

{
  "id": "233c1fa8-e6ea-4b4f-86b5-2bbccb39fd9e",
  "object": "chat.completion",
  "created": 1764635256,
  "model": "deepseek-reasoner",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Hello there! How can I assist you today?",
        "reasoning_content": "Hmm, the user just greeted me with "Hello!". This is a straightforward opening greeting, so I should respond warmly and invitingly. I can start with a cheerful "Hello there!" to mirror their friendly tone, then offer my assistance to encourage them to share what they need. Keeping it simple and open-ended works best here since they haven't specified a request yet."
      },
      "logprobs": null,
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 12,
    "completion_tokens": 109,
    "total_tokens": 121,
    "prompt_tokens_details": {
      "cached_tokens": 0
    },
    "completion_tokens_details": {
      "reasoning_tokens": 96
    },
    "prompt_cache_hit_tokens": 0,
    "prompt_cache_miss_tokens": 12
  },
  "system_fingerprint": "fp_eaab8d114b_prod0820_fp8_kvcache"
}

全部字段见“响应”一节

请求

curl -X POST 'https://api.iapp.co.th/v3/llm/deepseek-3p2/chat/completions' \
  -H 'apikey: YOUR_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{ \
    "model": "deepseek-reasoner", \
    "messages": [ \
      {"role": "user", "content": "สวัสดีครับ คุณช่วยอะไรได้บ้าง?"} \
    ], \
    "max_tokens": 4096, \
    "temperature": 0.7, \
    "top_p": 0.9 \
  }'

响应

200成功。字段在"响应"一节说明。

{
  "id": "233c1fa8-e6ea-4b4f-86b5-2bbccb39fd9e",
  "object": "chat.completion",
  "created": 1764635256,
  "model": "deepseek-reasoner",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Hello there! How can I assist you today?",
        "reasoning_content": "Hmm, the user just greeted me with "Hello!". This is a straightforward opening greeting, so I should respond warmly and invitingly. I can start with a cheerful "Hello there!" to mirror their friendly tone, then offer my assistance to encourage them to share what they need. Keeping it simple and open-ended works best here since they haven't specified a request yet."
      },
      "logprobs": null,
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 12,
    "completion_tokens": 109,
    "total_tokens": 121,
    "prompt_tokens_details": {
      "cached_tokens": 0
    },
    "completion_tokens_details": {
      "reasoning_tokens": 96
    },
    "prompt_cache_hit_tokens": 0,
    "prompt_cache_miss_tokens": 12
  },
  "system_fingerprint": "fp_eaab8d114b_prod0820_fp8_kvcache"
}

请求

JSON，采用 OpenAI 聊天补全的格式，OpenAI 客户端只需改 base URL 即可使用。

请求头

apikeystring必填

您的 API 密钥，在控制台的 API Keys 中创建。

请求体

modelstring必填

deepseek-reasoner（思考模式）或 deepseek-chat（更快的直接模式）。

messagesarray必填

对话：含 role（system、user 或 assistant）与 content 的对象。

max_tokensinteger

最多生成的词元数。默认 4096，最高 128,000。

streamboolean

设为 true 时以 server-sent events 逐词元返回。默认 false。

temperaturefloat

0 至 2，默认 0.7。代码和数学接近 0，分析用 1.0，对话和翻译用 1.3，创意写作用 1.5。

top_pfloat

核采样，默认 0.9。

价格

两个模型均为每 1K 输入词元 0.01 IC、每 1K 输出词元 0.02 IC。

响应

OpenAI 聊天补全对象，如右侧面板所示：答案在 choices[0].message.content，usage 为计费的词元数。使用 deepseek-reasoner 时，message.reasoning_content 为推理内容，usage.completion_tokens_details.reasoning_tokens 为其计数。stream: true 时以 data: 行逐块返回，每块含 choices[0].delta，以 data: [DONE] 结束。

OpenAI 聊天补全对象，如右侧面板所示：答案在 choices[0].message.content，usage 为计费的词元数。使用 deepseek-reasoner 时，message.reasoning_content 为推理内容，usage.completion_tokens_details.reasoning_tokens 为其计数。stream: true 时以 data: 行逐块返回，每块含 choices[0].delta，以 data: [DONE] 结束。

详情

实测

指标	结果
参数量	6850 亿
上下文	128,000 词元
数学与信息学奥赛	在 IMO 与 IOI 题目上达到金牌水平，据模型作者发布
许可	MIT；开放权重

限制

• 每个密钥每秒 5 次、每分钟 200 次请求。
• 推理词元按输出计费；设置 max_tokens 时请为其留出余量。