本文转载自 linux.do 社区干货帖，原帖需登录查看。作者亲自爬了 GitHub 1000+ 仓库按 star 排序，筛出仍在存活的免费渠道，每条标注了免费额度、是否绑卡、Base URL 和 Key 申请方式。

中文版

前言：网上那些”免费 xxx”帖子九成是过期/引流/已跑路，或者 0.001 倍 xxx 模型，大部分是降智或者映射的。这份是亲自爬了 GitHub 上 1000+ 仓库，按 star 排序筛出来还在活的，每条都标了：免费额度、要不要绑卡、Base URL、怎么申请 Key。直接抄作业就行。

一、永久免费层（注册就有，不绑卡）

• Google Gemini —— Base URL: https://generativelanguage.googleapis.com/v1beta；Key 申请: https://aistudio.google.com/app/apikey；免费额度: 15 RPM / 100 万 TPM / 1500 RPD；模型: Gemini 2.5 Flash / 2.5 Pro（限 RPD）；亮点: 1M 上下文，免费版最香；注意: 需 Google 账号，国内需魔法上网申请
• Groq —— Base URL: https://api.groq.com/openai/v1；免费额度: 30 RPM / 1.4 万 RPD；模型: Llama 3.3 70B / Qwen 2.5 / DeepSeek R1 Distill；亮点: 全球最快推理，几秒出结果；注意: 不绑卡，注册即用
• Cohere —— Base URL: https://api.cohere.com/v2；免费额度: 1000 次/月（trial key）；模型: Command R+ / Embed / Rerank；亮点: 自带 RAG 工具链，企业级；注意: trial key 60 天有效期，可重新申请
• Mistral AI —— Base URL: https://api.mistral.ai/v1；Key 申请: https://console.mistral.ai/api-keys；免费额度: 1M tokens / 两周（试用）；模型: Mistral Large / Codestral / Nemo；亮点: 欧洲公司，GDPR 合规；注意: 试用结束需绑卡继续
• Hugging Face Inference —— 免费额度: 按模型配额（每月重置）；模型: 3000+ 开源模型随便调；亮点: 什么模型都有，OpenAI 兼容；注意: 限速比较狠，重度用建议绑卡升级
• Cloudflare Workers AI —— Base URL: https://api.cloudflare.com/client/v4/accounts/{你的账号ID}/ai/run；Key 申请: https://dash.cloudflare.com/profile/api-tokens；免费额度: 1 万次/天 Neuron；模型: Llama 3 / Mistral / Qwen / gpt-oss；亮点: Cloudflare 免费层永不收费；注意: 需要 Cloudflare 账号

二、国内可用的免费 API（国内直连）

• DeepSeek —— Base URL: https://api.deepseek.com/v1；Key 申请: https://platform.deepseek.com/api_keys；免费额度: 注册送 ¥10；模型: deepseek-chat / deepseek-reasoner（R1）；亮点: 国内直连，送的钱够测试很久
• 智谱 GLM —— Base URL: https://open.bigmodel.cn/api/paas/v4；免费额度: GLM-4-Flash 永久免费，其余送 token；模型: GLM-4 / GLM-4V / GLM-4-Flash（免费）；亮点: GLM-4-Flash 永久免费这条很多人没注意
• 硅基流动 SiliconFlow —— Base URL: https://api.siliconflow.cn/v1；Key 申请: https://cloud.siliconflow.cn/account/ak；免费额度: 注册送 14 元 + 邀请送；模型: Qwen / DeepSeek / Llama / GLM 全家桶；亮点: 一个 Key 调所有国产开源模型，聚合最全
• 阿里通义 / 百炼 —— Base URL: https://dashscope.aliyun.com/compatible-mode/v1；Key 申请: https://bailian.console.alibabacloud.com；免费额度: Qwen-Turbo 永久免费（限 100 万 token/月）；模型: Qwen 全家桶；亮点: OpenAI 兼容，Cline / Cursor 直接接
• Kimi (Moonshot) —— Base URL: https://api.moonshot.cn/v1；Key 申请: https://platform.moonshot.cn/console/api-keys；免费额度: 注册送 15 元；模型: moonshot-v1-8k / 32k / 128k；亮点: 长上下文，128k 免费

三、聚合 API（一个 Key 调所有模型）

• OpenRouter —— Base URL: https://openrouter.ai/api/v1；免费额度: 充 $10 升 1K RPM；有几十个 :free 模型；模型: 全网聚合，标 :free 的随便调；亮点: 一个 Key 调遍全网，Cline/Cursor/LobeChat 通吃；注意: 免费模型有限速，重度用要充钱
• NVIDIA NIM —— Base URL: https://integrate.api.nvidia.com/v1；免费额度: 1000 次/月（每个模型）；模型: Nemotron / Llama / Qwen / DeepSeek；亮点: 4090 集群推理，速度快；注意: 需 NVIDIA 账号
• ModelScope —— Base URL: https://api-inference.modelscope.cn/v1；免费额度: 按模型配额；模型: Qwen / DeepSeek / GLM 等 60+；亮点: 阿里达摩院出品，国内直连
• Cerebras —— Base URL: https://api.cerebras.ai/v1；Key 申请: https://cloud.cerebras.ai；免费额度: 有限免费；模型: Llama 3.1 70B / 405B；亮点: WSE 芯片，比 Groq 还快

四、逆向 / 自部署方案（折腾党进）

• chatanywhere/GPT_API_free（44k star）—— GitHub: https://github.com/chatanywhere/GPT_API_free；Base URL: https://api.chatanywhere.tech（海外）/ https://api.chatanywhere.org（国内）；免费额度: 送 token，每日刷新，可调 GPT/Claude/Gemini/DeepSeek；亮点: 中转站，国内直连，价格低于官方；注意: 不是官方，有跑路风险，余额别充多
• gpt4free / gpt4free-ts（7k+ star）—— GitHub: https://github.com/xiangsx/gpt4free-ts；玩法: 自部署，逆向多家站点拿免费额度；亮点: 完全免费，OpenAI 兼容；注意: 逆向方案，稳定性看运气
• free-one-api —— GitHub: https://github.com/RockChinQ/free-one-api；玩法: 自部署网关，把 gpt4free/Bard/ChatGPT 转成 OpenAI 接口；Docker: docker run -d -p 3000:3000 rockchin/free-one-api；亮点: 一个面板管所有逆向源，统一 OpenAI 协议
• apfel（Mac 专属）—— GitHub: https://github.com/Arthur-Ficial/apfel；玩法: macOS 调用 Apple FoundationModels 本地模型；亮点: 完全本地推理，零成本零延迟，OpenAI 兼容；命令: brew install arthur-ficial/tap/apfel
• dsh-our-free-model（多源聚合）—— GitHub: https://github.com/Ebony-Vinyl/dsh-our-free-model；玩法: 自部署网关，聚合免费源；亮点: 不绑卡不注册，本地一键拉所有免费池

五、CLI / Coding Agent 免费方案

• GitHub Copilot —— 学生/开源维护者免费（education.github.com/pack），无限 Copilot + Copilot Chat，VSCode/IDE 原生集成
• Cursor / Cline 接免费源 —— Cline 选 OpenAI Compatible，Base URL 填上面任意一家；Cursor 在 Settings → Models → OpenAI API Base 填 Base URL + Key；推荐组合: Cline + 硅基流动 + DeepSeek V3（国内最快）、Cline + Groq + Llama 3.3 70B（国外最快）、Cline + Gemini 2.5 Pro（1M 上下文）
• Codex CLI（OpenAI 开源）—— GitHub: https://github.com/openai/codex；接任意 OpenAI 兼容 Base URL，配置 ~/.codex/config.toml

六、Prompt 工程免费资源

• awesome-chatgpt-prompts —— GitHub: https://github.com/f/awesome-chatgpt-prompts，100+ 角色提示词模板
• OpenAI 官方 Prompt 工程文档 —— OpenAI API 文档中的 Prompt engineering 章节
• DeepLearning.AI 短课 —— deeplearning.ai 的 ChatGPT Prompt Engineering for Developers，Andrew Ng 亲自讲

七、掏心窝建议

• 别想着白嫖干一切。免费层有配额，要么按 RPM 限要么按 RPD 限，别拿来跑批量任务。
• 多源切换：注册 Google + Groq + Mistral + Cohere 四个号，国内再开 DeepSeek + 智谱 + 硅基流动，凑齐 7 个免费源，日常用够了。
• 配 OpenRouter 做兜底，$10 充一次解锁 1K RPM，重度用这个最稳。
• 自部署 gpt4free 是双刃剑：能跑但易封，别放生产环境。
• 国内环境首选硅基流动 + DeepSeek 官方，速度和稳定性都好，还不用绑卡。
• 接 Cline / Cursor 用 OpenAI Compatible 模式，填 Base URL + Key 就行。
• 永久免费层只有 Google Gemini 和 Cloudflare Workers AI 两家真免费不绑卡，其他都是注册送 token。
• 省 token 的招：用 Cerebras/Groq 做快速迭代，用 Gemini 1M 做长上下文，用 OpenRouter 做兜底，按场景切换。
• 别折腾破解版客户端、别用别人分享的 Key，封号比省钱快。
• 真要重度用，老实充钱。DeepSeek 官方 $0.27/百万 token，比免费折腾的时间成本低多了。

English Version

Translated from a linux.do community guide. The author crawled 1000+ GitHub repos ranked by stars to filter out still-alive free channels, each labeled with free quota, card requirements, Base URL and how to get a Key.

Most “free xxx” posts online are expired, clickbait, dead services, or 0.001x-rate mapped models. This guide only lists channels verified alive, with free quota, card requirements, Base URL and Key application for each.

1. Permanent free tiers (no card required)

• Google Gemini — https://generativelanguage.googleapis.com/v1beta；Key: https://aistudio.google.com/app/apikey；Quota: 15 RPM / 1M TPM / 1500 RPD；Models: Gemini 2.5 Flash / 2.5 Pro (RPD-limited)；Highlight: 1M context, best free tier；Note: needs Google account
• Groq — https://api.groq.com/openai/v1；Quota: 30 RPM / 14K RPD；Models: Llama 3.3 70B / Qwen 2.5 / DeepSeek R1 Distill；Highlight: world’s fastest inference；Note: no card, register and go
• Cohere — https://api.cohere.com/v2；Quota: 1000 calls/month (trial key)；Models: Command R+ / Embed / Rerank；Highlight: built-in RAG toolchain；Note: trial key valid 60 days, renewable
• Mistral AI — https://api.mistral.ai/v1；Key: https://console.mistral.ai/api-keys；Quota: 1M tokens / two weeks (trial)；Models: Mistral Large / Codestral / Nemo；Highlight: EU company, GDPR compliant；Note: card required after trial
• Hugging Face Inference — Quota: per-model, resets monthly；Models: 3000+ open models；Highlight: everything is there, OpenAI-compatible；Note: aggressive rate limits
• Cloudflare Workers AI — https://api.cloudflare.com/client/v4/accounts/{id}/ai/run；Key: https://dash.cloudflare.com/profile/api-tokens；Quota: 10K Neurons/day；Models: Llama 3 / Mistral / Qwen / gpt-oss；Highlight: Cloudflare free tier never charges；Note: needs Cloudflare account

2. Free APIs reachable from China

• DeepSeek — https://api.deepseek.com/v1；Key: https://platform.deepseek.com/api_keys；Quota: ¥10 credit on signup；Models: deepseek-chat / deepseek-reasoner (R1)
• Zhipu GLM — https://open.bigmodel.cn/api/paas/v4；Quota: GLM-4-Flash free forever, plus gifted tokens；Models: GLM-4 / GLM-4V / GLM-4-Flash (free)
• SiliconFlow — https://api.siliconflow.cn/v1；Key: https://cloud.siliconflow.cn/account/ak；Quota: ¥14 on signup + referral bonuses；Models: Qwen / DeepSeek / Llama / GLM；Highlight: one key for all Chinese open models
• Alibaba Qwen / Bailian — https://dashscope.aliyun.com/compatible-mode/v1；Key: https://bailian.console.alibabacloud.com；Quota: Qwen-Turbo free forever (1M tokens/month)；Highlight: OpenAI-compatible, works with Cline / Cursor directly
• Kimi (Moonshot) — https://api.moonshot.cn/v1；Key: https://platform.moonshot.cn/console/api-keys；Quota: ¥15 on signup；Models: moonshot-v1-8k / 32k / 128k；Highlight: long context king

3. Aggregator APIs (one key for all models)

• OpenRouter — https://openrouter.ai/api/v1；Quota: top up $10 for 1K RPM；dozens of :free models；Highlight: one key for everything, works with Cline/Cursor/LobeChat；Note: free models are rate-limited
• NVIDIA NIM — https://integrate.api.nvidia.com/v1；Quota: 1000 calls/month per model；Models: Nemotron / Llama / Qwen / DeepSeek；Highlight: fast inference on GPU clusters；Note: needs NVIDIA account
• ModelScope — https://api-inference.modelscope.cn/v1；Quota: per-model；Models: 60+ incl. Qwen / DeepSeek / GLM；Highlight: by Alibaba DAMO, direct connection in China
• Cerebras — https://api.cerebras.ai/v1；Key: https://cloud.cerebras.ai；Models: Llama 3.1 70B / 405B；Highlight: WSE chips, even faster than Groq

4. Reverse-engineered / self-hosted options (tinkerers only)

• chatanywhere/GPT_API_free (44k stars) — https://github.com/chatanywhere/GPT_API_free；Endpoints: https://api.chatanywhere.tech (global) / https://api.chatanywhere.org (China)；Quota: gifted tokens, refreshed daily；covers GPT/Claude/Gemini/DeepSeek；Highlight: relay station, cheaper than official；Note: unofficial, risk of shutdown — don’t top up much
• gpt4free / gpt4free-ts (7k+ stars) — https://github.com/xiangsx/gpt4free-ts；Self-hosted, reverse-engineers multiple sites for free quota；fully free, OpenAI-compatible；stability varies
• free-one-api — https://github.com/RockChinQ/free-one-api；Self-hosted gateway converting gpt4free/Bard/ChatGPT into OpenAI API；Docker: docker run -d -p 3000:3000 rockchin/free-one-api
• apfel (Mac only) — https://github.com/Arthur-Ficial/apfel；Runs Apple FoundationModels locally on macOS；fully on-device, zero cost, OpenAI-compatible；brew install arthur-ficial/tap/apfel
• dsh-our-free-model — https://github.com/Ebony-Vinyl/dsh-our-free-model；Self-hosted gateway aggregating free sources；no signup, no card

5. Free options for CLI / coding agents

• GitHub Copilot — free for students/open-source maintainers (education.github.com/pack), unlimited Copilot + Copilot Chat
• Cursor / Cline with free sources — Cline: pick OpenAI Compatible, paste any Base URL above；Cursor: Settings → Models → OpenAI API Base；Recommended combos: Cline + SiliconFlow + DeepSeek V3 (fastest in China), Cline + Groq + Llama 3.3 70B (fastest abroad), Cline + Gemini 2.5 Pro (1M context)
• Codex CLI (open-source by OpenAI) — https://github.com/openai/codex；works with any OpenAI-compatible Base URL via ~/.codex/config.toml

6. Free prompt-engineering resources

• awesome-chatgpt-prompts — https://github.com/f/awesome-chatgpt-prompts，100+ role prompt templates
• OpenAI official prompt engineering docs
• DeepLearning.AI short course — ChatGPT Prompt Engineering for Developers, taught by Andrew Ng

7. Honest advice

• Don’t expect free tiers to do everything — quotas are RPM- or RPD-limited, not for batch jobs.
• Rotate multiple sources: Google + Groq + Mistral + Cohere accounts, plus DeepSeek + Zhipu + SiliconFlow in China — 7 free sources cover daily use.
• Use OpenRouter as fallback: one $10 top-up unlocks 1K RPM, most stable for heavy use.
• Self-hosted gpt4free is a double-edged sword: works but gets banned easily, never for production.
• In China, SiliconFlow + official DeepSeek give the best speed and stability, no card needed.
• For Cline / Cursor use OpenAI Compatible mode — just fill in Base URL + Key.
• Only Google Gemini and Cloudflare Workers AI are truly free forever without a card; the rest are signup credits.
• Save tokens by scenario: Cerebras/Groq for fast iteration, Gemini 1M for long context, OpenRouter as fallback.
• Don’t use cracked clients or other people’s shared keys — getting banned is faster than saving money.
• For heavy use, just pay. DeepSeek official at $0.27/M tokens costs less than the time spent chasing free quotas.

来源 / Source

• 原帖 / Original thread: https://linux.do/t/topic/3041274（linux.do 社区，需登录查看）