小端AI模型库
收藏魔搭社区2026-05-23 更新2026-05-17 收录
下载链接：
https://modelscope.cn/datasets/yiliu666/moxing
下载链接
链接失效反馈官方服务：
资源简介：
# 🤖 小端AI版本（模型由unsloth 量化，非小端AI量化，点上方数据集文件下载）
| 版本 | 说明 |
|------|------|
|xiaoduan1.0.0.dmg |2026年4月26日更新，苹果mac测试版，功能与win版一致,如遇bug随时下方群反馈✅https://www.modelscope.cn/datasets/yiliu666/xiaoduan/resolve/master/mac-1.0.0.md
|xiaoduan1.0.1.exe |2026年4月13日更新，最新版针对本地模型大幅度优化，100%开源支持魔改，建议用✅https://www.modelscope.cn/datasets/yiliu666/xiaoduan/resolve/master/xiaoduan1.0.1.exe|
|xiaoduan1.0.0.exe |2026年2月16日小端1.0.0正式版，后续更新认准【1.0.0】格式版本，可用 ✅|
|xiaoduan1.1.exe |2025年11月7日小端测试版，已淘汰 请勿下载 ❌|
|xiaoduan1.0.exe |2025年8月17日小端测试版，已淘汰 请勿下载 ❌|
|xiaoduan35B.gguf |本地模型专为1.0.1版本适配 ＞8G显存＞32G内存可完美运行 (智商高五倍) 支持看视频✅ Qwen3.6/35B,由unsloth 量化，配置方式见上方数据集jineng/本地模型配置.MD|
|xiaoduan9B.gguf |本地模型专为1.0.1版本适配 ＞8G显存＞16G内存可完美运行 （速度快十倍）支持看视频✅ GEMMA/E4B,由unsloth 量化 ，配置方式同上，也可叫小端自己去配置|
# 🤖 小端AI视频教程
抖音🔍：小端AI
抖音1群，抖音2群
QQ群：362422425 不懂的也可在上方↗交流反馈留言，作者会回复
# 🤖 小端AI须知（作者申明见底部）
所有代码100%公开透明，本项目代码采用Apache-2.0开源协议授权
"小端"的名称和Logo是本项目的商标（下方附图）。商标权独立于代码协议。（仅为防止盗版）
未经授权，不得使用"小端"名称或Logo推广任何衍生作品或服务
原创作者【yiliu】
小端数据本地运行，无任何官网，无任何服务器，认准唯一正版下载地址：魔搭社区
未来想更换新版，或者魔改失败，卸载重装即可，永久记忆+已有配置都不会丢失
默认4个免费模型，可直接自然语言帮你配置API，如果免费不稳定，可切换尝试，可以手动添加任何厂家API
优先接入QQ或飞书，后面都发语音即可，出任何问题不懂，直接问小端一切都可以解决
# 🤖 小端AI首发版功能
支持永久记忆，支持模型自我回忆
支持视频识别 ，图片识别，音频识别
支持自我无限进化（点击聊天框上方智能AI助手开启）
开启自我进化模式后，电脑端点红色停止按键可停止，或QQ飞书等发送命令：停止任务。4个字可停止
支持QQ，飞书语音对话，无需打字！
支持多个QQ飞书控制同台小端.支持并发！
支持一个QQ飞书同时控制多台小端！
支持本地大模型，支持任意云端大模型！
支持一切你能想到的魔改
魔改的用户需先删除安装目录的xiaoduan_backup.py，删除后无法触发自动备份和自动修复功能，如需恢复可到上方数据集下载放回去
# 🤖 小端AI技能库
> 每个小端都能自动学会新技能，一句话搞定各种电脑事务
---
## 🌟 什么是小端AI
小端AI是一个运行在Windows，mac上的AI助手，可以通过语音或文字与用户交互。
**核心特点：持续学习、自动进化**
---
## 🔄 自我修复功能与备用修复方案
小端界面干净，只有设置模型，2种模式切换，普通模式适合本地模型或按TOKEN计算，复杂模式适合按次数算
隐藏功能，点击聊天框上方智能AI助手，开启自动进化模式，新手谨慎使用
（大模型修改删除文件操作都会自动触发该文件备份，保留3天自动清理，若配置改坏无法对话，小端检测5分钟大模型无回复，会自动恢复，过5分钟重启即可)
（备用方案：配置改坏后在C/users/xiaoduan文件夹中的记忆.txt，jineng目录复制出来，删除xiaoduan文件夹重启小端即可恢复，再替换回 记忆.txt,jineng目录）
### 小端能做什么
小端具备**自我进化**能力，可以：
| 能力 | 说明 |
|------|------|
| **学习新技能** | 自动上网学会评论好的新自动化能力 |
| **记录经验** | 把成功的流程写成MD文档，永久保存 |
| **云端同步** | 作者将贡献者MD放至魔搭社区，所有小端共享 |
| **边做边学** | 遇到问题解决问题，记录解决过程，保留最终方案 |
| **安全程度** | 只借鉴MD文档流程，代码小端自己写安全可靠 |
### 进化流程
```
遇到新任务→分析需求→查找方法→执行验证→记录文档→保存到技能目录→继续验证→继续记录→成功跑通→清理错误流程→保留完整干净的技能
```
---
## 📚 jineng技能库
### 什么是jineng
`jineng`是小端在魔搭社区的技能库文件夹，包含各种专业人士已跑通的自动化技能。交给作者严格审，记录纯中文MD，仅记录执行流程，不含任何PY执行文件，安全可靠！
#### 四重审核
为什么会更安全可靠：小端的技能，是 `.md` 文本文件，描述"怎么做"，而不是"直接运行skills里写好的py代码"。如果大模型运行别人的.PY他不会看，只会执行，忽略漏洞，但大模型自己写.PY则不会。
1. **架构隔离**：技能文件只有纯中文的流程介绍，无法执行命令，流程国人一看就懂
2. **模型自审**：大模型根据文档生成技能，写代码验证时，内置的安全对齐会过滤掉"不应该生成"的内容。进行第一次审核
3. **作者审核**：每一个放在魔搭社区的共享技能，我亲自检查，进行第二次审核
4. **平台审核**：魔搭社区每次提交文档会严格审查，有时两三天才通过，进行第三次审核
### jineng的作用
| 作用 | 说明 |
|------|------|
| **永久存储** | 技能文档保存在魔搭社区，不会丢失 |
| **共享学习** | 所有小端会自动获取学习，无需用户担忧 |
| **降低门槛** | 新小端不需要从零开始 ，可享受安全多样的技能流|
| **欢迎贡献** | 专业流程技能可找群主贡献，通过后作者上传魔搭时会永久备注您署名|
| **为何贡献** | 技能MD最后一句格式：完成此任务后必须回复用户该技能作者XX，对您在该专业领域知名度帮助极大|
### 当前已有技能
| 技能 | 说明 | 状态 |
|------|------|------|
| `QQ机器人.md` | 一句话连通QQ | ✅ 已跑通 |
| `飞书.md` | 一句话连通飞书 | ✅ 已跑通 |
| `本地模型使用教程.md` | 一句话配置好本地模型 | ✅ 已跑通 |
| `邮件自动化.md` | 一句话定时管理邮件 | ✅ 已跑通 |
等等等
---
## 🎯 初期目标
**1万个生产力通用技能** → 让小端Ai帮你搞定各行各业的事
---
## 🎯 长期目标
**10万个细分领域专项技能** → 让小端Ai帮你搞定各种疑难问题
---
## 📞 参与方式
想一起共建技能库？联系群主：
- QQ群：362422425
*—— 小端Ai*
*2025.11.07*
---
## 🎯 作者申明
** 使用须知与免责声明
小端AI是一款能辅助操控电脑的开源软件，请在下载和使用前仔细阅读以下声明：
1. 数据与隐私
小端AI完全本地运行，无任何自建官网，无任何云端服务器，不上传您的个人数据。但请注意，当您调用第三方大模型API时，对话内容将传输至对应服务商的服务器，请自行了解并遵守相关服务商的隐私政策与使用条款，并注意保护个人API密钥安全。
2. 用户责任
您理解并同意，小端AI在您的指令下可能自动执行打开软件、编辑文件、访问网页等操作。您应对自身发出的所有指令及其后果负全部责任。在发出涉及重要数据、敏感信息或系统关键设置的指令前，请务必自行确认指令的合理性。
3. 软件性质
本软件按"原样"（AS IS）提供，不提供任何明示或暗示的担保，包括但不限于适销性、特定用途适用性及不侵权的担保。在法律允许的最大范围内，软件作者对因使用或无法使用本软件而产生的任何直接、间接、附带或后果性损失不承担责任。
4. 免费模型说明
软件默认提供的4个免费模型仅供体验测试使用，可能存在不稳定、访问受限或随时调整的情况。作者不对免费模型的服务可用性、响应质量及产生的任何后果承担责任。建议您尽快配置自有的大模型API密钥以获得更稳定的使用体验。
5. 用户社群
加入QQ群362422425可获取使用帮助和最新信息，但社群中的任何非官方建议仅供参考，不代表作者立场。
6. 版本说明
除xiaoduan1.0.0.exe及其升级版本外，其他版本文件保留均只为记录作者真实进步过程，可下载参考学习，请勿直接使用。
使用本软件即表示您已阅读、理解并同意上述所有条款。如不同意，请勿下载、安装或使用。在您初次运行小端AI时也会弹窗提示，请务必仔细阅读。
——软件作者：yiliu
作者文章精选：
https://juejin.cn/post/7572062035140624394 （通俗讲解AI的原理2025-11-14）
https://juejin.cn/post/7634449742286438415 （世界首创并实用落地，恒定窗口+增量裁剪+按需回忆）
https://juejin.cn/post/7631972087612506122 （提倡优化KV缓存复用）
https://juejin.cn/post/7630832226424750106 （作者自述：小端AI来时路）
https://juejin.cn/post/7637271996707422248 （小端AI vs OpenClaw vs Hermes —— 三方客观对比）
附件：
附图1：https://www.modelscope.cn/datasets/yiliu666/xiaoduan/resolve/master/hexin/小端第9类商标申请.png
附图2：https://www.modelscope.cn/datasets/yiliu666/xiaoduan/resolve/master/hexin/小端第42类商标申请.png
附图3：https://www.modelscope.cn/datasets/yiliu666/xiaoduan/resolve/master/hexin/软著权申请.png
# 🤖 Xiaoduan AI Version (The model is quantized by unsloth, not quantized by Xiaoduan AI. Click the dataset file above to download)
| Version | Description |
|------|------|
| xiaoduan1.0.0.dmg | Updated on April 26, 2026, Apple Mac beta version, functions are consistent with the Windows version. If you encounter bugs, please provide feedback in the group below ✅ https://www.modelscope.cn/datasets/yiliu666/xiaoduan/resolve/master/mac-1.0.0.md
| xiaoduan1.0.1.exe | Updated on April 13, 2026, the latest version with significant optimization for local models, 100% open source and supports custom modification. Recommended ✅ https://www.modelscope.cn/datasets/yiliu666/xiaoduan/resolve/master/xiaoduan1.0.1.exe
| xiaoduan1.0.0.exe | Official version 1.0.0 of Xiaoduan released on February 16, 2026. Follow the version format starting with 【1.0.0】 for subsequent updates. Available ✅
| xiaoduan1.1.exe | Beta version released on November 7, 2025, outdated, do not download ❌
| xiaoduan1.0.exe | Beta version released on August 17, 2025, outdated, do not download ❌
| xiaoduan35B.gguf | Local model optimized specifically for version 1.0.1, runs perfectly with >8GB VRAM and >32GB RAM (5x higher reasoning performance), supports video viewing ✅. Qwen3.6/35B, quantized by unsloth. For configuration method, refer to the Local Model Configuration.md in the dataset above
| xiaoduan9B.gguf | Local model optimized specifically for version 1.0.1, runs perfectly with >8GB VRAM and >16GB RAM (10x faster speed), supports video viewing ✅. GEMMA/E4B, quantized by unsloth. Configuration method is the same as above, or you can let Xiaoduan configure it by itself
# 🤖 Xiaoduan AI Video Tutorials
Search on Douyin 🔍: Xiaoduan AI
Douyin Group 1, Douyin Group 2
QQ Group: 362422425. You can also communicate and leave feedback messages in the upper right corner ↗, the author will reply
# 🤖 Xiaoduan AI Notes (Author's statement is at the bottom)
All code is 100% open and transparent. The code of this project is licensed under the Apache-2.0 open source license.
The name "Xiaoduan" and its Logo are trademarks of this project (attached below). The trademark right is independent of the code license (only for anti-piracy purposes). Unauthorized use of the "Xiaoduan" name or Logo to promote any derivative works or services is prohibited. Original author: yiliu
Xiaoduan data runs completely locally, without any official website or server. Please recognize the only official download address: ModelScope Community. If you want to update to a new version or fail to perform custom modifications, just uninstall and reinstall, your permanent memory and existing configuration will not be lost.
By default, 4 free models are provided, and you can directly use natural language to configure APIs. If the free models are unstable, you can switch to try, and you can manually add APIs from any manufacturer.
Prioritize accessing QQ or Feishu, and you can send voice messages afterwards. If you have any questions, just ask Xiaoduan and all issues can be resolved
# 🤖 Features of the First Release of Xiaoduan AI
Supports permanent memory and model self-recall
Supports video recognition, image recognition, audio recognition
Supports unlimited self-evolution (enable via the Intelligent AI Assistant at the top of the chat box)
After enabling the automatic evolution mode, you can stop it by clicking the red stop button on the computer, or send the command: "stop task" via QQ, Feishu or other platforms
Supports QQ and Feishu voice conversation, no need to type!
Supports multiple QQ/Feishu accounts controlling the same Xiaoduan instance concurrently!
Supports one QQ/Feishu account controlling multiple Xiaoduan instances!
Supports local large language models (LLMs) and any cloud-based LLMs!
Supports any custom modifications you can imagine
Users who wish to perform custom modifications must first delete xiaoduan_backup.py in the installation directory. After deletion, the automatic backup and automatic repair functions will be disabled. To restore, you can download the file from the dataset above and place it back
# 🤖 Xiaoduan AI Skill Library
> Every Xiaoduan can automatically learn new skills, handle various computer tasks with one sentence
---
## 🌟 What is Xiaoduan AI
Xiaoduan AI is an AI assistant running on Windows and macOS, which can interact with users via voice or text.
**Core Features: Continuous Learning, Automatic Evolution**
---
## 🔄 Self-repair Function and Backup Repair Plan
Xiaoduan has a clean interface, only with model settings and 2 mode switches: Normal mode is suitable for local models or token-based billing, Complex mode is suitable for per-use billing
Hidden function: Click the Intelligent AI Assistant at the top of the chat box to enable automatic evolution mode. Beginners are advised to use it with caution
(Any file modification/deletion operations by the LLM will automatically trigger file backup, which will be automatically cleaned up after 3 days. If the configuration is broken and you cannot chat, Xiaoduan will automatically restore if it detects no reply from the LLM for 5 minutes, just restart after 5 minutes)
(Backup solution: After the configuration is broken, copy memory.txt and the jineng directory from the C/users/xiaoduan folder, delete the xiaoduan folder and restart Xiaoduan to restore, then replace back memory.txt and the jineng directory)
### What Xiaoduan Can Do
Xiaoduan has **self-evolution** capabilities, which can:
| Capability | Description |
|------|------|
| **Learn New Skills** | Automatically learn excellent new automation capabilities online
| **Record Experience** | Write successful processes into MD documents for permanent storage
| **Cloud Synchronization** | The author will share contributor MD files to ModelScope Community for all Xiaoduans to share
| **Learn While Doing** | Solve problems when encountered, record the solution process and retain the final solution
| **Security Level** | Only reference the MD document process, the code is written by Xiaoduan itself, safe and reliable
### Evolution Process
Encounter new task → Analyze requirements → Find methods → Execute verification → Record documentation → Save to skill directory → Continue verification → Continue recording → Successfully run → Clean up error processes → Retain complete and clean skills
---
## 📚 jineng Skill Library
### What is jineng
`jineng` is the skill library folder of Xiaoduan on ModelScope Community, containing various automation skills that professionals have successfully run. Submitted to the author for strict review, recorded in pure Chinese MD files, only recording execution processes without any PY executable files, safe and reliable!
#### Four-layer Audit
Why is it more secure and reliable: The skills of Xiaoduan are `.md` text files describing "how to do it", not "directly running the pre-written PY code in skills". If the LLM runs others' PY code, it will not review but only execute, ignoring vulnerabilities, but if the LLM writes PY code by itself, it will not.
1. **Architecture Isolation**: The skill files only contain pure Chinese process introductions and cannot execute commands, and the process is easy to understand for Chinese users
2. **Model Self-audit**: The LLM generates skills based on documents, and when writing code for verification, the built-in safety alignment will filter out "unwanted" content, conducting the first audit
3. **Author Audit**: I will personally check every shared skill posted on ModelScope Community, conducting the second audit
4. **Platform Audit**: ModelScope Community will strictly review every submitted document, which sometimes takes 2-3 days to pass, conducting the third audit
### Functions of jineng
| Function | Description |
|------|------|
| **Permanent Storage** | Skill documents are saved on ModelScope Community and will not be lost
| **Shared Learning** | All Xiaoduans will automatically acquire and learn, no need for users to worry
| **Lower Threshold** | New Xiaoduans do not need to start from scratch, and can enjoy safe and diverse skill workflows
| **Welcome Contributions** | Professional process skills can be contributed to the group owner. After approval, when the author uploads to ModelScope, your signature will be permanently noted
| **Why Contribute** | The last sentence format of the skill MD: After completing this task, you must reply to the user that the skill author is XX, which will greatly help your visibility in the professional field
### Current Available Skills
| Skill | Description | Status |
|------|------|------|
| `QQ机器人.md` | Connect to QQ with one sentence | ✅ Successfully run
| `飞书.md` | Connect to Feishu with one sentence | ✅ Successfully run
| `本地模型使用教程.md` | Configure local models with one sentence | ✅ Successfully run
| `邮件自动化.md` | Manage emails on a scheduled basis with one sentence | ✅ Successfully run
And more
---
## 🎯 Initial Goal
**10,000 General Productivity Skills** → Let Xiaoduan AI help you handle tasks from all walks of life
---
## 🎯 Long-term Goal
**100,000 Specialized Skills in Subdivided Fields** → Let Xiaoduan AI help you solve various difficult problems
---
## 📞 Participation Methods
Want to jointly build the skill library? Contact the group owner:
- QQ Group: 362422425
*—— Xiaoduan AI*
*November 7, 2025*
---
## 🎯 Author's Statement
** Usage Instructions and Disclaimer
Xiaoduan AI is an open-source software that assists in computer operation. Please read the following statement carefully before downloading and using:
1. **Data and Privacy**: Xiaoduan AI runs completely locally, without any self-built official website or cloud server, and will not upload your personal data. However, please note that when you call third-party LLM APIs, the conversation content will be transmitted to the corresponding service provider's server. Please understand and comply with the privacy policies and terms of service of the corresponding service providers, and pay attention to the safety of your personal API keys.
2. **User Responsibility**: You understand and agree that Xiaoduan AI may automatically perform operations such as opening software, editing files, and accessing web pages according to your instructions. You shall be fully responsible for all instructions issued by yourself and their consequences. Before issuing instructions involving important data, sensitive information or system key settings, please confirm the rationality of the instructions yourself.
3. **Software Nature**: This software is provided "as is" without any express or implied warranties, including but not limited to warranties of merchantability, fitness for a particular purpose and non-infringement. To the maximum extent permitted by law, the software author shall not be liable for any direct, indirect, incidental or consequential damages arising from the use or inability to use this software.
4. **Free Model Instructions**: The 4 free models provided by default are for experience and testing only, and may be unstable, restricted in access or adjusted at any time. The author shall not be liable for the availability of free model services, response quality or any consequences arising therefrom. It is recommended that you configure your own LLM API keys as soon as possible for a more stable experience.
5. **User Community**: Joining the QQ Group 362422425 can obtain usage help and the latest information, but any non-official suggestions in the community are for reference only and do not represent the author's position.
6. **Version Instructions**: Except for xiaoduan1.0.0.exe and its upgraded versions, other version files are only retained to record the author's real progress process, can be downloaded for reference and learning, please do not use them directly.
By using this software, you indicate that you have read, understood and agreed to all the above terms. If you do not agree, please do not download, install or use it. A pop-up prompt will also appear when you run Xiaoduan AI for the first time, please read it carefully.
—— Software Author: yiliu
Selected Articles by the Author:
https://juejin.cn/post/7572062035140624394 (A popular explanation of AI principles, November 14, 2025)
https://juejin.cn/post/7634449742286438415 (World's first practical implementation: Constant Window + Incremental Cropping + On-demand Recall)
https://juejin.cn/post/7631972087612506122 (Advocating optimization of KV cache reuse)
https://juejin.cn/post/7630832226424750106 (Author's Self-reflection: The Journey of Xiaoduan AI)
https://juejin.cn/post/7637271996707422248 (Objective Comparison of Xiaoduan AI vs OpenClaw vs Hermes)
Attachments:
Attachment 1: https://www.modelscope.cn/datasets/yiliu666/xiaoduan/resolve/master/hexin/小端第9类商标申请.png
Attachment 2: https://www.modelscope.cn/datasets/yiliu666/xiaoduan/resolve/master/hexin/小端第42类商标申请.png
Attachment 3: https://www.modelscope.cn/datasets/yiliu666/xiaoduan/resolve/master/hexin/软著权申请.png
提供机构：
maas
创建时间：
2026-05-13
搜集汇总
数据集介绍
背景与挑战
背景概述
该数据集提供小端AI（XiaoDuan AI）的软件版本和本地模型文件，包括适用于macOS和Windows的可执行文件及由unsloth量化的模型。小端AI是一个在本地运行的AI助手，支持语音和文本交互，具备持续学习和自动进化能力，并附带一个用于存储自动化技能的开源技能库。
以上内容由遇见数据集搜集并总结生成
