本文来源:https://mp.weixin.qq.com/s/S0-DUKyr1AhSxz-tKdMMRA
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
FMR v2.0.1发布,兼容图片生成、多模态输入自动路由、能力分池调度,安装更稳、路由更聪明。
从"只能聊天"到"能文能图",FMR 正式进入多模态时代。
一、FMR是什么:一个可自动扩展的个人免费Token池
如果你每个月的AI账单烧掉几百甚至几千块,如果你在不同的免费API渠道之间来回切换、注册、填Key、记额度,如果你因为某个免费模型突然下架或限额而中断工作——FreeModelRouter(FMR)就是为你做的。
FMR 的核心理念很简单:把全网分散的免费模型资源,聚合成一个像付费API一样稳定、统一的个人Token池。
它兼容 OpenAI 和 Anthropic 两种主流接口协议,理论上兼容所有 Agent 框架接入——涵盖 OpenClaw、Hermes Agent、Claude Code、QClaw、WorkBuddy 等。你不需要改代码,只需要把 API 端点指向本地运行的 FMR,剩下的交给它。
社区网址(必看,内含指引): https://community.freemodel.eu.org/
核心价值就两点:
•自动扩展Token池:全网搜集免费模型资源,新渠道上线自动接入,旧渠道下线自动切换
•智能调度算法:让不稳定的免费资源,像付费订阅一样稳定流畅
二、v2.0.1 核心更新:从文本到多模态的跨越
v2.0.1 是 FMR 的一次架构级升级。底层路由引擎全面重构,正式兼容多模态能力。以下是值得关注的6项核心更新:
1. 兼容图片生成:免费Token也能画图了
这是 v2.0.1 最重磅的新能力。
FMR 现在兼容 OpenAI Images API (POST /v1/images/generations),你可以直接借助 FMR 调用免费渠道的图片生成能力,无需关心背后是哪个 provider 在给出服务。
兼容的调用方式:
• 同步返回图片(provider 直接返回图片数据)
• 异步轮询(provider 返回任务ID,FMR 自动轮询直到完成)
• URL 和 base64 两种格式自动转换(FMR 会按你的需要格式返回)
这意味着什么? 以前想生图,要么掏钱买 Midjourney/DALL-E,要么逐个去免费渠道试。现在把 FMR 指向你的工具,直接调用就行——和调用文本模型一样简单。
2. 多模态输入自动路由:带图的请求找对的模型
难题情形: 你发了一张截图让 AI 分析,结果请求被路由到了一个纯文本模型,直接报错或返回空内容。
v2.0.1 的应对思路: 自动检测消息中是否包含图片内容,如果是,自动路由到兼容 vision(视觉理解)能力的模型。
兼容的检测范围:
• OpenAI 格式的 image_url 类型消息
• Anthropic 格式的 image 类型消息(涵盖 base64 和 URL 两种来源)
你什么都不用改,FMR 自动帮你把"带图的请求"送到"能看图的人"手里。
3. 能力路由引擎:让合适的模型做合适的事
v2.0.1 对路由引擎做了底层重构,引入了"能力分池"机制。
以前的难题: 所有请求都走同一个主模型和备用模型队列。一个模型可能在文本上很强,但根本不兼容图片——如果它被设为主模型,所有请求都会先砸向它,涵盖图片请求。
v2.0.1 的改进: 模型按能力分类,每个能力有独立的"热池":
| 文本池 | |
| 视觉池 | |
| 生图池 |
FMR 会依据请求的类型,自动从对应的能力池中挑选最佳模型。文本请求不会误打到生图模型,生图请求也不会被当成文本处理。
同时,模型列表 (/v1/models) 现在会暴露三种虚拟模型,让你的 Agent 工具能直观看到 FMR 兼容哪些能力:
•fmr-text — 纯文本对话
•fmr-vision — 多模态理解
•fmr-image-gen — 图片生成
4. 智能路由策略:均衡还是粘性,你说了算
v2.0.1 新增了两套路由策略,让调度更聪明:
BalancedStrategy(均衡策略): 将请求均匀分散到多个 provider,避免流量集中打到单一渠道导致限流。合适高并发、追求稳定性的情形。
StickyStrategy(粘性策略): 同一使用者的同一类请求优先走同一个 provider。好处是减少上下文切换——如果你和某个模型的对话有上下文记忆,粘性策略会让后续请求尽量走同一个渠道,感受更连贯。
策略在后台自动运行,你不需要手动规格。
5. 精细化超时规格:不同任务给不同耐心
不同能力的请求,合理等待时间本来就不同:
• 文本对话:几秒就该有响应
• 图片生成:等个十几秒很正常
• 复杂推理:可能需要更久
v2.0.1 兼容按能力、按 provider 规格超时时间,不再一刀切。FMR 会依据请求类型自动应用合理的超时预算,减少"明明还在处理却被误判为失败"的情况。
6. 安装更稳、启动更顺
v2.0.1 对安装和启动流程做了多项优化:
•Node.js 二进制包直装:安装脚本现在兼容直接下载预编译的 Node.js 二进制包,不用等源码编译,安装速度大幅提高
•ZIP 解压和权限检测优化:安装过程中自动检测解压是否完整、权限是否正确,减少"装完跑不起来"的难题
•启动预检 + 错误诊断:启动脚本增加了前置检查,常见难题(端口占用、规格缺失、网络不通)会给出明确的诊断提示
•Provider 自动恢复:之前因为临时故障被标记为不可用的 provider,现在会自动尝试恢复,不用手动重启
•稳如老狗:真实长上下文任务实测,连续数小时工作,几百次调用,几千万免费token消耗无中断。
三、安装与使用:三分钟上手
三种安装方式
Windows 使用者:下载安装包,双击运行,一路 Next 即可
Linux / macOS 使用者:打开终端,粘贴下面这行命令,回车
curl -sSL https://freemodel.eu.org/install.sh | bash OpenClaw 使用者:把下面这条指令发送给 OpenClaw,它会帮你完成全部安装
openclaw skills install free-model-router-v1 三步上手
1.下载安装(1分钟)
2.打开管理台填入 API Key(2分钟):浏览器自动打开 http://127.0.0.1:5678/admin,按页面提示复制粘贴你的免费API Key。不知道去哪拿Key?页面下方有各渠道的注册指南。
3.开始使用:安装完成后,FMR 在本地 http://127.0.0.1:5678/v1 运行。如果你用 Claude Code、Cursor 或其他兼容自定义API端点的工具,把 Provider 地址指向这个本地端点即可。OpenClaw 使用者安装时会自动完成这一步。
图片生成怎么用
如果你使用的工具兼容 OpenAI Images API(如部分 Agent 框架或自定义脚本),把 API 端点指向 FMR:
POST http://127.0.0.1:5678/v1/images/generations 参数和 OpenAI 官方 API 一致,FMR 会自动帮你路由到兼容图片生成的免费渠道。
四、为什么值得信任
🔒 你的对话从不上报 FMR 是本地代理。你的提示词、图片和模型响应只经过你的电脑和 Provider 服务器,运营方不收集、不存储任何内容。
📊 运营完全透明 社区公告借助管理台「事件通知」平台同步推送,不是黑箱运营。所有模型可用性数据来自社区成员自愿上报(默认关闭,随时可关)。
🚀 更新节奏公开 每两周一个迭代版本,路线图公开在文档站。你的反馈会直接进 GitHub issue,可被追踪、被回复。
五、写在最后
v2.0.1 的发布,标志着 FMR 从"文本专用"走向"多模态全能"。
图片生成能力的加入,意味着你可以用免费 Token 做更多事——不只是聊天和写代码,还可以生成配图、外观草图、视觉素材。多模态输入的自动路由,让你发一张图问"这什么意思"时,不用再手动切换模型。
能力路由引擎的重构,是底层架构的一次质变。它让 FMR 能更好地适应免费模型生态的多样性:有的模型擅长文本,有的能看图,有的会生图——FMR 让每种能力都找到最合适的 provider。
免费模型随时可能下架,但社区永远在线。
如果你还没试过 FMR,现在就是最好的时机。访问 社区网站 了解更多,扫码加入交流群,与走在时代前沿的AI实践者同行。
本文部分机型信息由 FreeModelRouter 开发团队给出,能力描述基于 v2.0.1 版本实测。
This article comes from: https://mp.weixin.qq.com/s/S0-DUKyr1AhSxz-tKdMMRA. If there are any copyright issues, please contact us at: 13951431913#139.com or submit a comment on the right.
FMR v2.0.1 released, compatible with image generation, multi-modal input auto-routing, and capability-based pool scheduling, making installation more stable and routing smarter.
From "chat-only" to "text-and-image capable", FMR has officially entered the multi-modal era.
1. What is FMR: A self-expanding personal free token pool
If your monthly AI bill eats up hundreds or even thousands of dollars, if you're constantly switching between different free API channels to register, paste keys, and track quotas, or if your work is interrupted because a free model suddenly gets shut down or rate-limited — FreeModelRouter (FMR) is built for you.
The core idea of FMR is simple: aggregate scattered free model resources across the web into a single, stable, and unified personal token pool that behaves like a paid API.
It is compatible with the two mainstream interfaces — OpenAI and Anthropic — and theoretically supports integration with all agent frameworks, including OpenClaw, Hermes Agent, Claude Code, QClaw, WorkBuddy, and more. You don't need to change your code; just point the API endpoint to your local FMR instance and let it handle the rest.
Community website (must-read, contains guides): https://community.freemodel.eu.org/
The core value boils down to two points:
• Automatic token pool expansion: continuously scan for new free model resources online; add new channels automatically as they appear, and switch away from deprecated channels.
• Intelligent scheduling algorithms: make unstable free resources behave as reliably as a paid subscription.
2. Core updates in v2.0.1: from text to multi-modal
v2.0.1 represents a significant architectural upgrade for FMR. The underlying routing engine has been fully rebuilt to natively support multi-modal capabilities. Here are six key updates worth noting:
1. Image generation compatibility: use free tokens to draw pictures
This is the most exciting new feature in v2.0.1.
FMR now supports the OpenAI Images API (POST /v1/images/generations), allowing you to invoke image generation from free channels via FMR without worrying about which provider is serving the image.
Compatible invocation methods include:
• Synchronous image return (provider sends image data directly)
• Asynchronous polling (provider returns a task ID; FMR polls until completion)
• Automatic conversion between URL and base64 formats (FMR formats the output as requested)
What does this mean? Previously, generating images required either paying for Midjourney/DALL-E or manually trying out free channels. Now, with FMR configured in your tool, you can generate images as easily as calling a text model.
2. Multi-modal input auto-routing: send images to the correct model
The problem: You send a screenshot for analysis, but the request gets routed to a text-only model, resulting in an error or empty response.
The v2.0.1 solution: Automatically detect whether a message contains an image. If it does, route the request to a vision-capable model.
This detection covers:
• OpenAI-style image_url messages
• Anthropic-style image messages (both base64 and URL sources)
You don't need to modify your setup; FMR automatically sends "requests with pictures" to "models that can read pictures".
3. Capability-based routing engine: the right model for the right job
v2.0.1 introduces a "capability pooling" mechanism, fundamentally changing how models are selected.
The old issue: All requests went through a single primary model or a fallback queue. If the primary model couldn't handle images, all image requests failed immediately.
The new approach: Models are categorized by capability, with separate "hot pools" for:
Capability pool
Task types handled
Text-only conversations, code completions
Multi-modal understanding (image + text), visual analysis
Image generation
FMR selects the best model from the appropriate pool based on the request type. Text requests won't be sent to image generators, and image requests won't be sent to text-only models.
Additionally, the /v1/models list now exposes three virtual models to clearly indicate which capabilities are supported:
• fmr-text — pure text conversation
• fmr-vision — multi-modal understanding
• fmr-image-gen — image generation
4. Smarter routing strategies: balance vs. stickiness
v2.0.1 adds two new routing strategies for more intelligent load management:
BalancedStrategy: Distributes requests evenly across providers to prevent overloading a single channel. This is ideal for high-concurrency applications that prioritize reliability.
StickyStrategy: Assigns the same user's requests of the same type to the same provider whenever possible. This reduces context switching and preserves conversation history, which is crucial for maintaining state in long-running dialogues.
The strategies operate automatically in the background, requiring no manual intervention.
5. Granular timeout configuration: patience where it matters
Different tasks require different amounts of time.
• Text conversations should respond within seconds.
• Image generation can take tens of seconds.
• Complex reasoning might need even more time.
v2.0.1 allows for capacity-specific and provider-specific timeouts. FMR automatically applies the appropriate maximum wait time based on the request type, reducing false negatives where a model is simply taking longer than the default threshold.
6. Smoother installation and startup
v2.0.1 includes several quality-of-life improvements for getting started:
• Binary Node.js installation: The installer now downloads a pre-compiled Node.js executable, eliminating the need for manual compilation and significantly speeding up setup.
• Improved ZIP extraction and permission checks: The installation process now verifies that files have been extracted correctly and that the necessary permissions are set, reducing "cannot run" errors.
• Startup diagnostics and error messages: The launcher performs preliminary checks for common issues (e.g., port conflicts, missing configurations, network connectivity) and provides clear error messages.
• Provider recovery: If a provider fails, FMR marks it as unavailable but automatically attempts to recover it later without requiring a full restart.
• Proven stability: In real-world tests with long-context tasks, FMR remained operational for hours, handling hundreds of calls and consuming tens of millions of free tokens without interruption.
3. Installation and Usage: Get started in three minutes
Three installation methods
Windows users: Download the installer, double-click, and follow the prompts.
Linux/macOS users: Open a terminal and paste the following command:
OpenClaw users: Send this command to your OpenClaw instance, and it will handle the entire installation process for you.
Three simple steps to get running
1. Download and install (1 minute)
2. Open the admin panel and enter your API keys (2 minutes): The browser will automatically open http://127.0.0.1:5678/admin. Follow the on-screen instructions to copy and paste your free API keys. If you're unsure where to find them, the page provides links to the registration guides for each service.
3. Start using FMR: Once installed, FMR runs locally at http://127.0.0.1:5678/v1. If you use tools like Claude Code, Cursor, or any other application that supports custom API endpoints, simply point the provider address to this local endpoint. OpenClaw users will have this configured automatically during the install command.
How to use image generation
If your tool supports the OpenAI Images API (such as certain agent frameworks or custom scripts), configure the API endpoint to point to FMR:
The parameters are identical to the official OpenAI API. FMR will automatically route your request to a compatible free image generation service.
4. Why FMR is trustworthy
🔒 Your conversations are private: FMR is a local proxy. All data—prompts, images, and responses—travel directly between your computer and the provider's server. No data is collected, stored, or logged by the FMR developers or operators.
📊 Transparent operations: Community announcements are distributed via the in-app "Event Notification" system rather than external black-box channels. All model availability data is sourced from voluntary user reports (opt-in).
🚀 Open development roadmap: FMR follows a strict bi-weekly release schedule. The roadmap is publicly available on the documentation site, and all user feedback is tracked transparently in the GitHub issue tracker.
5. Final thoughts
The release of v2.0.1 marks a significant evolution for FMR, transitioning it from a "text-only" service to a "multi-modal" powerhouse.
The addition of image generation capability expands the utility of your free tokens. You can now generate visual assets, such as plot illustrations or concept sketches, alongside your text-based analysis. The automatic routing for multi-modal inputs ensures that sending a picture with a question yields a meaningful analysis rather than an error.
The reconstruction of the capability routing engine represents a fundamental improvement in how FMR adapts to the free model ecosystem. In a landscape where some models excel at vision and others at text, FMR ensures that the most appropriate model is selected for every task.
While free models are always subject to change (shutdowns, rate limits), the FMR community is stable and continuously improving. Visit the community website to learn more, join the discussion group, and collaborate with fellow AI enthusiasts to make the most of these free resources.
Portions of this model information were provided by the FreeModelRouter development team, and the capability descriptions are based on actual testing of the v2.0.1 release.
Some of the model information in this article is provided by the FreeModelRouter development team, and the capability description is based on actual testing of version 2.0.1.