国内大模型通义千问、DeepSeek 等大多做了OpenAI 兼容接口可以直接用openaiSDK只改api_key、base_url、model三个参数即可。下面给你可直接跑的完整示例。一、公共准备通用安装官方 SDK无需装其他包bash运行pip install -U openai python-dotenv新建.env存密钥不要硬写在代码里env# 通义千问 QWEN_API_KEYsk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx QWEN_BASE_URLhttps://dashscope.aliyuncs.com/compatible-mode/v1 # DeepSeek DEEPSEEK_API_KEYsk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx DEEPSEEK_BASE_URLhttps://api.deepseek.com二、调用通义千问Qwen1基础调用python运行import os from dotenv import load_dotenv from openai import OpenAI load_dotenv() client OpenAI( api_keyos.getenv(QWEN_API_KEY), base_urlos.getenv(QWEN_BASE_URL) ) resp client.chat.completions.create( modelqwen-max, # 可选qwen-plus、qwen-turbo、qwen3-max 等 messages[ {role: system, content: 你是 helpful 助手}, {role: user, content: 简要介绍 ROS2} ], temperature0.7, max_tokens1024 ) print(resp.choices[0].message.content)2流式输出逐字返回python运行stream client.chat.completions.create( modelqwen-plus, messages[{role: user, content: 写一段快速排序 Python 代码}], streamTrue ) for chunk in stream: if chunk.choices[0].delta.content: print(chunk.choices[0].delta.content, end, flushTrue) print()三、调用 DeepSeek1基础调用python运行import os from dotenv import load_dotenv from openai import OpenAI load_dotenv() client OpenAI( api_keyos.getenv(DEEPSEEK_API_KEY), base_urlos.getenv(DEEPSEEK_BASE_URL) ) resp client.chat.completions.create( modeldeepseek-chat, # 或 deepseek-reasoner思考模式 messages[ {role: system, content: 你是专业技术顾问}, {role: user, content: 解释一下大模型的 SFT 与 RLHF} ], temperature0.7 ) print(resp.choices[0].message.content)2流式输出python运行stream client.chat.completions.create( modeldeepseek-chat, messages[{role: user, content: 用 5 句话总结深度学习}], streamTrue ) for chunk in stream: if chunk.choices[0].delta.content: print(chunk.choices[0].delta.content, end, flushTrue) print()四、关键参数对照表一眼看懂表格模型base_url常用 model获取 key 地址通义千问https://dashscope.aliyuncs.com/compatible-mode/v1qwen-max、qwen-plus、qwen-turbo阿里云百炼控制台DeepSeekhttps://api.deepseek.comdeepseek-chat、deepseek-reasonerDeepSeek 官网控制台五、常见坑 解决401 鉴权失败检查 api_key 是否正确、是否带sk-前缀确认该 key 已开通对应模型权限404 路径不存在千问必须带/compatible-mode/v1DeepSeek/v1可加可不加官方示例用根域名超时 / 网络慢千问优先用国内dashscope.aliyuncs.comDeepSeek 国内访问不稳时可考虑合规中转模型名报错千问不要写gpt-3.5必须写qwen-xxxDeepSeek用deepseek-chat/deepseek-reasoner六、多模型统一切换模板推荐python运行def get_client(provider: str) - OpenAI: load_dotenv() if provider qwen: return OpenAI(api_keyos.getenv(QWEN_API_KEY), base_urlos.getenv(QWEN_BASE_URL)) elif provider deepseek: return OpenAI(api_keyos.getenv(DEEPSEEK_API_KEY), base_urlos.getenv(DEEPSEEK_BASE_URL)) else: raise ValueError(unsupported provider) # 切换只需改这里 client get_client(qwen) # 或 deepseek
