Token太贵?这个项目对接了15+免费大模型,智能路由无感切换,每月13亿token免费用Tired of high token costs? This project connects to 15+ free AI models, with intelligent routing for seamless switching, offering up to 1.3 billion free tokens per month.


AI free token/AI免费积分

本文来源:https://mp.weixin.qq.com/s/SV-xv0bNxUM9jtI7jQJRCg
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

封面

AI Agent的火爆正在带动大模型 token 集体涨价。

智谱 GLM 的 Coding Plan 每天 10:00 放量,几分钟售罄,年内已经三度提价累计 83%。腾讯 CodeBuddy 企业版价格直接翻倍。字节豆包上线专业版收费,加强版 200 元/月,5 小时额度没做多少任务就烧完。

这是整个行业的转向——从"烧钱抢用户"到"让用户买单"。40块一个月挺划算?不是幻觉,是之前各平台在补贴。2026 年中,补贴结束了。

但有一个事实容易被忽略——说"AI太贵"的前提是你只用一家。如果愿意花十几分钟注册几个免费账号,把各家免费额度统一调度呢?每个大模型厂商其实都在提供免费额度。 Google 每天 1500 次免费推理、Groq 每天 7000 次、GitHub Models 每月 $150 额度、OpenRouter 和 HuggingFace 也有免费套餐……

单独看每一个都不够用,但 15+ 家叠在一起,就是每月约 13 亿 Token 的免费池子。

免费额度池示意

但问题在于每次注册一个新平台就要换一个 SDK、换一个 Key,还要写代码做负载均衡。

15+家免费模型聚合

有个开源项目叫 FreeLLMAPI,就是做这件事的。它解决的不是注册,是注册完之后的统一调度。


这是怎么做到的

FreeLLMAPI 本质是一个智能路由器。它把 15+ 家厂商的免费 API 全部接进来,统一成一个 /v1 接口。

统一调度示意你的代码连这一个地址,发请求时指定 model="auto",路由器自己决定用哪个模型。

它聚合了哪些模型

模型列表截图

项目聚合了 15+ 家厂商的免费模型。以下是部分示例(完整列表见项目官网):

提供商代表模型说明
GoogleGemini 2.5 Flash / 3.x有免费额度,速率限制较低
GroqLlama 4 / Qwen 3速度快,免费额度充足
GitHub ModelsGPT-4.1 / GPT-4o月度免费额度
MistralLarge 3 / Codestral每日免费额度
CloudflareKimi K2 / GLM-4.7免费,低延迟
HuggingFaceDeepSeek V4 / Qwen 3每日免费推理
OpenRouter21 个免费模型聚合多个来源
智谱 (Z.ai)GLM-4.7 Flash国内模型,无需代理

还有 OpenAI 兼容的自定义提供商——你可以把自己的付费 API Key 也加进去,允许路由器优先用免费额度,免费额度用完自动走你的付费 Key。

六种路由策略

六种路由策略

路由器不只"谁有空用谁"。六种模式对应不同场景:

auto 是默认选项,适合大多数人。路由器每几秒检测一次各模型的响应延时和剩余额度,谁快用谁。你写一封邮件它可能选 Gemini,写一段代码它选 Llama,做一个翻译它选 GPT-4o-mini——不需要你手动干预。

priority 让你指定顺序。比如你先想用智谱(国内快),智谱不行走 Groq,Groq 不行走 Google。路由器按顺序尝试,前面的不行自动跳到后面的。

smartest 和 fastest 是一对。写方案、做分析时用 smartest 选最强模型(GPT-4o 或 Gemini),简单问答用 fastest 选轻量模型(Gemini Flash 或 Groq 上的小模型)。可以手动切,也可以让路由器根据 prompt 长度自动判断。

reliable 适合对稳定性有要求的场景。路由器会避开近期错误率高的提供商,选最不容易报错的。

custom 是高级模式,你可以定义路由链规则——比如"当 prompt 超过 2000 tokens 时走 Gemini,否则走 Groq"。通过管理面板配置。

故障转移和粘性会话

如果正在用的模型突然报 429 限速或 5xx 错误,路由器会自动切换到下一个可用的模型。切换过程你是无感的——不会看到报错,只会收到其他可用大模型的回复。

同一个话题如果中途换模型,AI 会丢失刚才的上下文。FreeLLMAPI 会自动保持同一个模型 30 分钟,只有当你开启新话题或超时才尝试切换。

额度预警

管理面板会显示每个提供商今天的用量、剩余额度、响应速度排行。你可以直观地看到哪个模型今天快用完了,提前手动切换策略——而不是等到请求报错了才知道。

用量统计面板
装一个服务,一劳永逸

怎么装

先用好桌面版。 去 GitHub Releases 下载对应系统的安装包(Windows .exe 或 macOS .dmg),安装后在系统托盘运行。

但使用之前需要先注册几个免费账户。 FreeLLMAPI 本身不生产免费额度,它是把你已注册的各家厂商免费额度统一调度。以下是你需要提前注册的平台:

平台注册难度说明
OpenRouter低邮箱注册,免费模型多;充值 $10 后日限额从 50 提升到 1000
Google AI Studio低有谷歌账号即可,免费额度充足
Groq低邮箱注册,无需海外卡,速度最快
GitHub Models低有 GitHub 账号即可,每月免费额度
HuggingFace低邮箱注册,无需海外卡
智谱 (Z.ai)国内直连手机号注册,无需代理

注册后在控制台生成 API Key,然后在 FreeLLMAPI 管理面板的 Keys 页面逐一添加。

代码中只需要改一行(统一 API Key 在管理面板 Keys 页面获取):

from openai import OpenAIclient = OpenAI(    base_url="http://localhost:3001/v1",    api_key="freellmapi-your-key",)

其他一切不变。model="auto" 让路由器自动选模型,也可以指定具体模型名。

你也可以混入自己的付费 Key(比如 OpenAI 的付费 API),路由器会自动优先用免费额度,免费的用完再走你的付费 Key,不会浪费一分钱。

如果想用 Docker 部署或从头编译,项目 GitHub 上有完整的安装文档。


什么时候该用,什么时候不该用

适合的场景:

Agent日常写作、AI 聊天、内容创作、原型验证、作业汇报、个人项目——这些场景模型偶有切换不影响最终结果。你省下的可能是一年几千块的 API 费用。

不适合的场景:

代码补全和 IDE 内联建议——模型切换会打乱补全逻辑,前一秒用 GPT-4o 后一秒切到 Gemini,上下文不一致会导致补全质量跳水。

生产环境——需要确定性的模型行为、固定的延迟表现、可追溯的调用链路。免费额度的不确定性和路由器的黑盒决策不适合线上业务。

一句话判断: 如果 AI 输出的质量波动你能接受,用它省钱;如果需求更精准的输出质量,就搭一个固定的付费Plan。


觉得有用的话,点个关注+赞再走~我是专注AI自动化的启明,咱们下一篇接着聊。

Source: https://mp.weixin.qq.com/s/SV-xv0bNxUM9jtI7jQJRCg If there are any copyright issues, please contact: 13951431913#139.com or submit a comment on the right.

The popularity of AI Agents is driving the collective price increase of large model tokens.

Zhipu GLM's Coding Plan releases capacity at 10:00 every day and sells out within minutes, with prices rising by 83% in three increases this year. Tencent CodeBuddy's enterprise version price has doubled directly. ByteDance Doubao has launched a paid professional version, with a strengthened version costing 200 yuan/month, and the 5-hour quota is exhausted after completing very few tasks.

This is a shift across the entire industry — from "burning money to grab users" to "making users pay". Is 40 yuan per month a good deal? It's not an illusion; the platforms were subsidizing before. As of mid-2026, the subsidies have ended.

But a fact that is easy to overlook is that saying "AI is too expensive" assumes you only use one service. What if you're willing to spend ten minutes registering several free accounts to consolidate the free quotas? In reality, every large model vendor offers free quotas. Google provides 1500 free inferences daily, Groq 7000, GitHub Models $150 per month, and OpenRouter and HuggingFace also have free packages...

Individually, none of them are sufficient for heavy use, but pooled together, they form a free pool of roughly 1.3 billion tokens per month.

The problem is that registering a new platform requires changing an SDK or Key each time, which complicates the process.

There is an open-source project called FreeLLMAPI that addresses this issue. It doesn't assist with registration; rather, it provides a unified dispatch for the free quotas you've already registered for.

How does it work?

FreeLLMAPI is essentially an intelligent router. It aggregates the free APIs of 15+ vendors into a single /v1 interface.

Your code connects to this address, and when sending a request, it specifies model="auto", allowing the router to decide which model to use.统一调度示意

Which models are aggregated?

The project aggregates free models from 15+ vendors. Examples include:

Additionally, there are custom providers compatible with OpenAI — you can add your own paid API Key, and the router will prioritize using the free quota, falling back to your paid key once the free quota is exhausted.

Six routing strategies

The router is not limited to "use who's available". Six modes cater to different scenarios:

The auto mode is the default, suitable for most people. The router checks the response latency and remaining quota of each model every few seconds, selecting the fastest one. For example, it might choose Gemini for an email, Llama for code, and GPT-4o-mini for a translation — all without manual intervention.

The priority mode allows you to specify an order. For instance, you can prefer Zhipu (for domestic speed), then Groq, and finally Google. The router will try them in sequence, skipping over unavailable ones.

The smartest and fastest modes are a pair. Use smartest for high-quality tasks like analysis (e.g., GPT-4o or Gemini), and fastest for simple questions (e.g., Gemini Flash or a small Groq model). The router can automatically select based on the prompt length.

The reliable mode is for high-availability needs, where the router avoids providers with recent error spikes and selects the most robust option.

The custom mode is for advanced users who need to define specific routing rules — for example, using Gemini for prompts over 2000 tokens and Groq for shorter ones.

Failover and Sticky Sessions

If the current model returns a 429 (rate limit) or 5xx error, the router automatically switches to the next available model. This transition is seamless — you won't see an error message, just the response from a different model.

To maintain context in a conversation, the router enforces "sticky sessions" — it will continue using the same model for 30 minutes or until the topic changes, ensuring consistency in a single dialogue.

Quota Alerts

The management panel displays the daily usage and remaining quota for each provider, along with a speed ranking. This allows you to visually identify which models are nearing their limits and switch strategies proactively rather than reacting to errors.

Getting Started

First, install and run the desktop version. Download the appropriate installer (Windows .exe or macOS .dmg) from the GitHub releases page. Once installed, it runs in the system tray.

Before using the service, you must register for free accounts with the various providers, as FreeLLMAPI does not create the quotas — it merely aggregates them. Platforms to register include:

After generating an API Key from each provider's console, head to the Keys page in the FreeLLMAPI management panel to add them one by one.

In your code, only one line needs to be changed — point it to the unified API provided by the application:

The rest of your configuration remains unchanged. Using model="auto" lets the router select the best model, while specific model names can be used for direct selection.

You can also mix in your own paid keys (e.g., for OpenAI). The router will prioritize free quotas first, only using the paid key when the free allowance is exhausted.

For users interested in more deployment options, such as running the service locally via Docker or compiling from source, the GitHub repository provides comprehensive installation instructions.

When to Use, and When Not to Use

Suitable Scenarios:

Ideal for everyday tasks such as writing, chatting, content creation, prototyping, homework assignments, and personal projects where the occasional shift in model quality is acceptable. This setup can potentially save thousands of dollars in annual API costs.

Unsuitable Scenarios:

Avoid using it for code completion or IDE-based suggestions, as the abrupt switching of models can lead to inconsistent context and poor completion quality.

Similarly, it is not recommended for production environments that require deterministic behavior, constant latency, and traceable call paths, as the free quota dynamics and routing logic introduce too much uncertainty for reliable business operations.

In summary, if you can tolerate minor fluctuations in output quality to save money, this tool is for you. If high precision and stability are non-negotiable, stick to a dedicated paid subscription plan.

If you find this helpful, please leave a like and subscribe — I'm Qiming, dedicated to AI automation, and we'll continue the discussion in the next article.

封面

免费额度池示意

15+家免费模型聚合

模型列表截图

六种路由策略

用量统计面板

装一个服务,一劳永逸

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry