[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"pack-detail-china-ai-api-alternatives-zh":3,"seo:pack:china-ai-api-alternatives:zh":104},{"code":4,"message":5,"data":6},200,"操作成功",{"pack":7},{"slug":8,"icon":9,"tone":10,"status":11,"status_label":12,"title":13,"description":14,"items":15,"install_cmd":103},"china-ai-api-alternatives","🐲","#0EA5E9","new","本周新建","国内 AI API 平替 — 不出国、不用 VPN、人民币付费的 OpenAI \u002F Claude 替代","十个国内 AI API 与网关：DeepSeek-V3 \u002F R1 \u002F Coder、Qwen、ChatGLM、Kimi（Moonshot）、MiniMax，加 One API \u002F LiteLLM Proxy \u002F Cherry Studio 做路由和兜底。人民币结算、可开票、兼容 OpenAI SDK、无需 VPN。",[16,28,35,42,52,60,68,78,86,95],{"id":17,"uuid":18,"slug":19,"title":20,"description":21,"author_name":22,"view_count":23,"vote_count":24,"lang_type":25,"type":26,"type_label":27},2832,"1b0d1ab2-1edb-49e1-9853-b02807a64140","deepseek-v3-open-weight-671b-moe-model-with-gpt-4o-quality","DeepSeek-V3 — Open-Weight 671B MoE Model with GPT-4o Quality","DeepSeek-V3 is a 671B-param MoE model (37B active per token). Matches GPT-4o on benchmarks. MIT-licensed weights, $0.27\u002F1M input on the hosted API.","DeepSeek",62,0,"en","skill","Skill",{"id":29,"uuid":30,"slug":31,"title":32,"description":33,"author_name":22,"view_count":34,"vote_count":24,"lang_type":25,"type":26,"type_label":27},2833,"c8ffbe43-1354-4034-8c86-6b0ab3076998","deepseek-r1-open-weight-reasoning-model-rivaling-openai-o1","DeepSeek-R1 — Open-Weight Reasoning Model Rivaling OpenAI o1","DeepSeek-R1 is the open-weight reasoning model that matches OpenAI o1 on math, code, science benchmarks. Streaming chain-of-thought visible. MIT-licensed.",64,{"id":36,"uuid":37,"slug":38,"title":39,"description":40,"author_name":22,"view_count":41,"vote_count":24,"lang_type":25,"type":26,"type_label":27},2834,"08acf3a7-b56b-40d2-9c94-9a8eb773eca4","deepseek-coder-code-specialized-model-for-local-inference","DeepSeek Coder — Code-Specialized Model for Local Inference","DeepSeek Coder is the code-specialized open-weight model with FIM (fill-in-middle) support. Beats Codestral on HumanEval. Drops into Continue, Aider.",67,{"id":43,"uuid":44,"slug":45,"title":46,"description":47,"author_name":48,"view_count":49,"vote_count":24,"lang_type":25,"type":50,"type_label":51},3022,"2dff2d41-ba09-4f17-bcce-22b9537c1db9","qwen-code-terminal-coding-agent-for-qwen-models","Qwen Code — Terminal Coding Agent for Qwen Models","Qwen Code is an open-source terminal coding agent for Qwen models. Node 22+, npm or Homebrew install, \u002Fauth flow, codebase Q&A, refactors, and tests.","QwenLM",97,"script","Script",{"id":53,"uuid":54,"slug":55,"title":56,"description":57,"author_name":58,"view_count":59,"vote_count":24,"lang_type":25,"type":26,"type_label":27},2264,"98bef1e7-42b9-11f1-9bc6-00163e2b0d79","chatglm-open-bilingual-chat-model-tsinghua-keg-98bef1e7","ChatGLM — Open Bilingual Chat Model by Tsinghua KEG","ChatGLM is a family of open bilingual language models from Tsinghua University that support English and Chinese conversation, code generation, and tool use, with variants optimized for consumer GPU deployment.","Script Depot",146,{"id":61,"uuid":62,"slug":63,"title":64,"description":65,"author_name":66,"view_count":67,"vote_count":24,"lang_type":25,"type":50,"type_label":51},3643,"b6547ac2-21f2-5e33-98db-4ddc3552d152","oh-my-kimi-evidence-gated-agent-runtime-for-kimi","oh-my-kimi — Evidence-gated Agent Runtime for Kimi","oh-my-kimi (OMK) is a CLI runtime that adds evidence gates and worktree isolation to Kimi Code; verified 69★ and ships `omk init\u002Fdoctor\u002Fchat`.","Skill Factory",103,{"id":69,"uuid":70,"slug":71,"title":72,"description":73,"author_name":74,"view_count":75,"vote_count":24,"lang_type":25,"type":76,"type_label":77},3932,"aadf52c3-ffd0-5a2f-8e04-95de3bb54dfb","minimax-mcp-official-minimax-mcp-server","MiniMax-MCP — Official MiniMax MCP Server","Official MiniMax Model Context Protocol server exposing media-generation tools (audio, image, video, music); verified 1474★, pushed 2026-05-14.","MCP Hub",92,"mcp","MCP",{"id":79,"uuid":80,"slug":81,"title":82,"description":83,"author_name":84,"view_count":85,"vote_count":24,"lang_type":25,"type":26,"type_label":27},3821,"65f4b9a7-d932-51b8-aa22-6f0ff7cff276","one-api-unified-llm-api-gateway-docker","One API — Unified LLM API Gateway (Docker)","One API is a self-hosted LLM API gateway: unify OpenAI\u002FClaude\u002FGemini\u002FDeepSeek endpoints, manage keys, and deploy via Docker in minutes (33.7k★).","AI Open Source",89,{"id":87,"uuid":88,"slug":89,"title":90,"description":91,"author_name":92,"view_count":75,"vote_count":24,"lang_type":25,"type":93,"type_label":94},2789,"0f113965-1adc-4435-982b-fb613fa4d157","litellm-proxy-unified-gateway-for-100-llm-apis","LiteLLM Proxy — Unified Gateway for 100+ LLM APIs","LiteLLM Proxy maps 100+ LLM providers (Anthropic, OpenAI, Bedrock, Vertex) to one OpenAI-compatible endpoint. Auth, rate limit, cost track, fallbacks.","LiteLLM (BerriAI)","agent","Agent",{"id":96,"uuid":97,"slug":98,"title":99,"description":100,"author_name":101,"view_count":102,"vote_count":24,"lang_type":25,"type":26,"type_label":27},2821,"121c2d1e-678a-4ea4-8799-70391f0a11de","cherry-studio-custom-models-byok-any-llm-provider","Cherry Studio Custom Models — BYOK Any LLM Provider","Cherry Studio Custom Models adds any OpenAI-compatible endpoint — proxy, local, or third-party. Mix Claude, GPT, Gemini, DeepSeek, Ollama side-by-side.","Cherry Studio",107,"tokrepo install pack\u002Fchina-ai-api-alternatives",{"pageType":105,"pageKey":8,"locale":106,"title":107,"metaDescription":108,"h1":109,"tldr":110,"bodyMarkdown":111,"faq":112,"schema":128,"internalLinks":135,"citations":148,"wordCount":161,"generatedAt":162},"pack","zh","国内 AI API 平替 — 10 个不出国 \u002F 不用 VPN \u002F 人民币付费的 OpenAI \u002F Claude 替代","DeepSeek-V3 \u002F R1 \u002F Coder、Qwen、ChatGLM、Kimi（Moonshot）、MiniMax — 加 One API \u002F LiteLLM Proxy \u002F Cherry Studio 做路由。十个国内 AI API 与网关:人民币结算、可开发票、兼容 OpenAI SDK、不需要 VPN,企业合规也能过。","国内 AI API 平替 — OpenAI \u002F Claude 替代选型","面向想用上 GPT-4o \u002F Claude \u002F o1 质量但**不办海外卡、不挂 VPN、要开发票**的国内团队,挑了十个:DeepSeek-V3 做通用便宜底座、DeepSeek-R1 做 o1 级推理、DeepSeek Coder 做代码、Qwen Code 和 ChatGLM 做双语兜底、Kimi 做长文档、MiniMax 做语音多模态;One API 做国内能跑的 OpenAI 兼容网关、LiteLLM Proxy 做多厂商失败转移、Cherry Studio 做发版前的桌面 sanity check。全部支持人民币结算、主流厂商都能开发票、对 OpenAI SDK 调用 drop-in 替换。","## 这个 pack 解决什么\n\n国内团队上 AI 功能,每次都撞同样三堵墙:(1) OpenAI \u002F Anthropic 注册流程要海外卡 + 海外 IP — 个人项目能蒙,公司层面非常痛。(2) 在国内裸调 api.openai.com \u002F api.anthropic.com 不稳定,必须挂代理,而企业代理本身又有合规问题。(3) 财务要**增值税专票**,海外 SaaS 不开。\n\n本 pack 选 10 个资产把这三堵墙一次性补齐,且**不牺牲 GPT-4o 级质量**。分三档:**前沿档(DeepSeek-V3 \u002F R1 \u002F Coder)**、**国产厂商替代档(Qwen、ChatGLM、Kimi、MiniMax)**、**把厂商选择藏在一个 OpenAI 兼容端点后面的网关档(One API、LiteLLM Proxy、Cherry Studio)**。装完你的代码基本不变 — 只改 `base_url` 和 `model` — 财务每月都能拿到发票。\n\n## 推荐安装顺序 — 先免费 \u002F 便宜,再上质量,最后接路由\n\n1. **DeepSeek-V3 — 671B MoE 开源大模型,GPT-4o 级质量**(id 2832)。从这里起步。托管 API 每百万输入 token 约 $0.27 — 比 GPT-4o 便宜 10 倍 — 在 platform.deepseek.com 用人民币结算。OpenAI 兼容:`base_url` 切 `https:\u002F\u002Fapi.deepseek.com\u002Fv1`、`model` 设 `deepseek-chat`、发版。权重 MIT 开源,将来想自托管随时切。\n2. **Qwen Code — 通义千问的终端编码 Agent**(id 3022)。阿里云 DashScope 托管 Qwen2.5 \u002F Qwen3,OpenAI 兼容端点、人民币结算。Qwen Code 是大多数国内开发者用来把 Qwen 接入编码循环的 CLI 壳子,不用自己写 SDK 胶水。API 和这个 CLI 第一天就一起装。\n3. **ChatGLM — 清华 KEG 开源双语聊天模型**(id 2264)。智谱 \u002F 清华那条线的经典。bigmodel.cn 上的 GLM-4 系列人民币结算、支持发票,是非 DeepSeek 选择里最强的**双语客户面向**工作负载选项 — 翻译、客服、中英混合内容审核。\n4. **oh-my-kimi — Kimi 的证据门控 Agent 运行时**(id 3643)。Moonshot 的 Kimi 是国内**长上下文**冠军(200K+ token、文档级理解)。处理\"把整份合同 \u002F 招标书 \u002F 手册扔进去\"这类 GPT-4o 算不过账的工作。oh-my-kimi 在外面套一层 agent runtime,让模型不只是个聊天框。\n5. **MiniMax-MCP — MiniMax 官方 MCP 服务器**(id 3932)。MiniMax 强在语音和多模态(TTS、声音克隆、视频转文本)。官方 MCP 服务器给任何支持 MCP 的客户端(Claude Desktop \u002F Cursor \u002F Cline)一个工具表面来调这些模型。需要文本以外能力时再装。\n6. **DeepSeek-R1 — 对标 OpenAI o1 的开源推理模型**(id 2833)。V3 的推理兄弟。同一套 API 形状,不同 model 别名(`deepseek-reasoner`)。**硬推理**调用走这里、其他都走 V3 — 第 8 步的网关让这只是一行配置。\n7. **DeepSeek Coder — 代码专用模型,本地推理**(id 2834)。更小、专做代码的分支 — 工作站 GPU 就能跑,把代码补全 \u002F IDE 内重构的网络跳直接砍掉。**放开发者电脑上,不放生产网关里**。\n8. **One API — 统一 LLM API 网关(Docker)**(id 3821)。本 pack 里最重要的一装。One API 是开源、国内能用的**OpenAI 兼容网关** — 它讲 OpenAI 的请求 \u002F 响应格式,把调用路由给 DeepSeek、Qwen、ChatGLM、Kimi、MiniMax、Azure、Anthropic 等几十家。Docker 跑在你自己的 VPC 里,代码指过去,**换厂商变成一行配置而不是一次代码改动**。下游厂商的人民币结算不受影响。\n9. **LiteLLM Proxy — 100+ LLM API 统一网关**(id 2789)。西方做的等价物。如果团队是 Python 栈、想要开箱即用的 per-key 成本追踪、或者要更成熟的 failover 和限流逻辑,用这个替代 One API。形状一样 — 一个 OpenAI 兼容端点,后面挂多厂商。\n10. **Cherry Studio Custom Models — BYOK 接任何 LLM 厂商**(id 2821)。桌面客户端。在你把任何一项接进生产之前,先把 key 粘到 Cherry Studio 里,在**真实 prompt 上**验证模型回答的方式符合你应用的预期。这是在你 commit 路由决策之前**对比 DeepSeek vs Qwen vs Kimi** 最便宜的方法。\n\n## 它们怎么协同\n\n```\n          [ 前沿档 ]                     [ 厂商替代档 ]                  [ 本地兜底 ]\n          DeepSeek-V3(通用)             Qwen(双语 \u002F 阿里系)            DeepSeek Coder(笔记本)\n          DeepSeek-R1(推理)             ChatGLM(智谱 \u002F 清华)\n                  │                      Kimi(长上下文 \u002F Moonshot)\n                  │                      MiniMax(语音 \u002F 多模态)\n                  │                              │\n                  └────────── One API ───────────┤\n                              (Docker、OpenAI 兼容)\n                                       │\n                              LiteLLM Proxy(备选路由层)\n                                       │\n                                 你的应用代码\n                                       │\n                              Cherry Studio(开发期肉眼验证)\n```\n\n主干是 **DeepSeek-V3 + One API + LiteLLM Proxy + Cherry Studio**。这四件套覆盖通用工作负载、把厂商选择藏在一个端点后面、给你一个带成本追踪的路由层、给一个桌面客户端做肉眼校对。Qwen \u002F ChatGLM \u002F Kimi \u002F MiniMax \u002F DeepSeek-R1 \u002F Coder 都是**网关后面的备选 backend**,按工作负载需要的特长加进来。\n\n## 上线前要先想清楚的取舍\n\n- **上下文长度。** Kimi 领先(200K+),Qwen 和 DeepSeek 多数套餐 64K-128K,GLM-4 约 128K。工作负载是\"总结 80 页 PDF\"时,Kimi 是国内**唯一不用分块**的答案。\n- **中英能力差异。** 五家都能讲英语。DeepSeek 和 Qwen 在英文 benchmark 上最接近 GPT-4o;GLM 和 Kimi 在创意英语上稍弱、在中文上反超。**用你自己的 prompt 测**,benchmark 均值不预测你的工作负载。\n- **计价单位。** 多数厂商在 dashboard 上按人民币百万 token 报价,在公开 API 页又给一个美元价。两个数字不一定一对一 — 有时候人民币更便宜,有时候不。**永远把用量写进自己的账本**,对账不要只信 dashboard。\n- **限速和并发。** 前沿模型新账号通常起步是 60 RPM \u002F 1 并发。生产工作负载要申请提额 — 有时候要打电话。**留一周提前量**,不要在能上 10 RPS 的功能上线当天才申请。\n- **推理模型的延迟。** R1 \u002F GLM-4-Reasoning \u002F Qwen-QwQ 拿速度换质量。一次推理调用 30-90 秒 — 必须接流式响应或异步队列,**不要让它阻塞同步请求**。\n\n## 常见踩坑\n\n- **把厂商 SDK 写死。** 全场 `from openai import OpenAI` 然后只切 `base_url`,你保留了可移植性。直接 `import dashscope` \u002F `import zhipuai`,**你就和一家厂商绑死了**、走不了网关。所有厂商都公开 OpenAI 兼容端点 — 用 OpenAI SDK 调它们。\n- **把 ICP \u002F 备案当 API 问题。** 备案是给**国内托管的对公网网站**用的、不是给 API 调用用的。**境外服务器调 DeepSeek API 不需要备案。**当你的域名是 `*.cn` \u002F 服务国内用户时**才**需要 ICP。这两件事分开看。\n- **忘记 token 计数不通用。** OpenAI 的 tokenization(tiktoken \u002F cl100k)和 DeepSeek \u002F Qwen \u002F GLM 不一样。1000 字中文 prompt 在 OpenAI 算 350 token,在 DeepSeek 可能是 280。**用 tiktoken 估的成本会差 10-30%** — 要做预算时永远用厂商自己的 tokenizer。\n- **拿生产 key 直接打生产端点测试。** 本 pack 里每家厂商都有 sandbox \u002F 测试 key,**用它**。否则你 CI 第一次卡死循环,就是 Qwen-Max 给你出一张 200 块的意外账单。\n- **跳过发票流程。** 主流厂商(DeepSeek \u002F 阿里云 \u002F 智谱 \u002F Moonshot \u002F MiniMax)都支持发票,但每家流程都不一样。**注册当周就让财务提交申请** — 第一张发票要 30 天周期,过期不补开。\n- **以为 `deepseek-chat` 永远是 V3。** 厂商别名会变。`deepseek-chat` 指向当下的聊天旗舰;今天是 V3、明天可能是 V3.5 \u002F V4。如果你的 eval 依赖**某个具体权重**,**钉死具体版本字符串**、不要靠别名。\n",[113,116,119,122,125],{"q":114,"a":115},"DeepSeek vs Qwen 怎么选?","**英文偏重的成本敏感工作负载选 DeepSeek-V3、阿里系或双语客户面向工作负载选 Qwen。** DeepSeek 的 API 是当下达到 GPT-4o 级质量最便宜的路径(约 $0.27 \u002F 1M 输入 token),权重 MIT 开源、想自托管随时切。Qwen 的优势在**集成深度** — 它原生跑在阿里云 DashScope 上、和你其他云资源同一套账户、双语客户面向 UI 的回答质量最强、Qwen Code CLI 让你不用写 SDK 胶水就有完整编码循环。**API only + 英文重 → 先装 DeepSeek。已经在阿里云付计算 \u002F 存储 + 想一家厂商搞定一切 → 先装 Qwen。**",{"q":117,"a":118},"API 消费能开增值税专票(发票)吗?","**主流厂商都能开**,但流程不一样。DeepSeek 在 platform.deepseek.com 账单页按月开 — 公司税号填一次、每月点\"申请发票\"。阿里云(Qwen)走标准阿里云发票通道 — 账户认证成**企业账号**之后全自动。智谱(ChatGLM)和 Moonshot(Kimi)首次需要给财务团队发邮件提供公司信息,之后每月自动开。MiniMax 同。**第一笔消费到第一张发票留 30 天滞后** — 厂商按月后开。",{"q":120,"a":121},"企业合规要看哪几点?","**五件事。** (1) **数据主权** — 确认 API 端点是境内托管;DeepSeek \u002F 阿里云 DashScope \u002F 智谱 \u002F Moonshot \u002F MiniMax **全在境内**。(2) **Prompt \u002F 响应日志** — 每家厂商隐私政策都说**可能记录调用用于安全审查**;工作负载含 PII 或商业秘密时,要么发前脱敏、要么谈一个 no-log 企业套餐。(3) **跨境传输** — 应用把非国内用户的数据发给境内 API 时,PIPL 跨境规则适用,要问法务。(4) **算法备案** — 在国内上对公网的生成式 AI 功能时,网信办要求**算法备案**;这是**你的**责任、不是模型厂商的。(5) **发票 + 合同** — 确保你签合同的主体和开发票的主体**对得上**,对不上财务对账就头疼。",{"q":123,"a":124},"响应速度怎么样?","**从国内网络看,五家国产厂商对 OpenAI \u002F Anthropic 是降维打击** — 往返完全在国内、跳过国际链路。DeepSeek-V3 和 Qwen 聊天模型首 token 通常 200-500ms 出;GLM-4 和 Kimi 相近。**推理模型**(DeepSeek-R1 \u002F Qwen-QwQ \u002F GLM-4-Reasoning)慢 — 30-90 秒一次回答,因为它要先生成内部思维链才答 — 这是**模型种类的固有属性**、不是中国 vs 海外问题。和国内裸调 OpenAI \u002F Anthropic 对比,国产厂商**不止更快,是真的能调通**。",{"q":126,"a":127},"能直接用 OpenAI SDK 接吗?还是要重写?","**保留 OpenAI SDK。** 本 pack 里每家厂商都公开 OpenAI 兼容端点,前面再套 One API \u002F LiteLLM Proxy,就给所有厂商一个统一的 OpenAI 兼容端点。实战:装 OpenAI SDK、`base_url` 指你的网关、`model` 设网关路由到的那个(`deepseek-chat` \u002F `qwen-max` \u002F `glm-4` \u002F `moonshot-v1-128k` 等),剩下的 — `chat.completions.create` \u002F 流式 \u002F tool use \u002F JSON 模式 — **一行不动**。OpenAI 兼容层**唯一覆盖不了**的是 OpenAI 独有的前沿功能(Realtime API \u002F Assistants API \u002F GPT 图像编辑)。**95% 的 LLM 工作负载,SDK 替换就是一行配置。**",{"@context":129,"@type":130,"name":131,"description":132,"numberOfItems":133,"inLanguage":134},"https:\u002F\u002Fschema.org","ItemList","国内 AI API 平替","十个国内 AI API 与网关,按推荐安装顺序排,给想用上 GPT-4o \u002F Claude \u002F o1 质量但不办海外卡、不挂 VPN、要开发票的国内团队用 — 涵盖 DeepSeek-V3 \u002F R1 \u002F Coder、Qwen、ChatGLM、Kimi(Moonshot)、MiniMax、One API、LiteLLM Proxy、Cherry Studio。",10,"zh-CN",[136,140,144],{"url":137,"anchor":138,"reason":139},"\u002Fzh\u002Ftopics\u002Fai-cost-optimization-stack","AI 成本优化栈 pack","兄弟 pack,讲单次调用的成本天花板和路由技巧,和本 pack 的网关档配套",{"url":141,"anchor":142,"reason":143},"\u002Fzh\u002Ftopics\u002Fmcp-search-rag-tools","MCP 搜索 + RAG 工具 pack","国产 API 接通之后下一道缺口是公网搜索 + 私有文档检索,那个 pack 接着补",{"url":145,"anchor":146,"reason":147},"\u002Fzh\u002Fai-tools-for\u002Fapi-gateway","TokRepo API 网关分类","本 pack 之外的更多 OpenAI 兼容网关和代理",[149,153,157],{"claim":150,"source_name":151,"source_url":152},"DeepSeek 公开 OpenAI 兼容 API 端点 api.deepseek.com\u002Fv1","DeepSeek API 文档","https:\u002F\u002Fapi-docs.deepseek.com\u002F",{"claim":154,"source_name":155,"source_url":156},"阿里云 DashScope 通过 OpenAI 兼容端点暴露 Qwen 系列模型","阿里云 DashScope 文档","https:\u002F\u002Fhelp.aliyun.com\u002Fzh\u002Fdashscope\u002F",{"claim":158,"source_name":159,"source_url":160},"One API 是开源的 OpenAI 兼容多厂商 LLM 网关","One API GitHub 仓库","https:\u002F\u002Fgithub.com\u002Fsongquanpeng\u002Fone-api",1980,"2026-05-23T00:00:00Z"]