> For the complete documentation index, see [llms.txt](https://unsloth.ai/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://unsloth.ai/docs/zh/docs.md).

# Unsloth 文档

Unsloth 让你通过开源界面在自己的本地硬件上运行并训练 AI 模型。

我们的文档将指导你在本地运行并训练自己的 LLM。

<a href="https://unsloth.ai/download" class="button primary" data-icon="down-to-bracket">下载 Unsloth</a><a href="/pages/7fc6e99d6b528309cd47374238dd9e205e0676a8#features" class="button secondary" data-icon="sparkles">功能</a><a href="https://github.com/unslothai/unsloth" class="button secondary" data-icon="github">GitHub</a>

<table data-card-size="large" data-view="cards"><thead><tr><th></th><th></th><th data-hidden data-card-cover data-type="image">封面图</th><th data-hidden data-card-target data-type="content-ref"></th></tr></thead><tbody><tr><td><h4>隆重推出 Unsloth Desktop</h4></td><td>首个用于运行和训练 LLM 以及扩散模型的桌面应用。</td><td><a href="/files/2f6d683fff2b297edbdcc5c1043bc7a986309170">/files/2f6d683fff2b297edbdcc5c1043bc7a986309170</a></td><td><a href="/pages/7fc6e99d6b528309cd47374238dd9e205e0676a8">/pages/7fc6e99d6b528309cd47374238dd9e205e0676a8</a></td></tr><tr><td><h4>Qwen3.8-27B 来了！</h4></td><td>在 Unsloth 中运行 Qwen3.8-27B 和 Qwen3.8-2.4T！</td><td><a href="/files/0fb600e9cfe1c255e855383e83372687ada48fff">/files/0fb600e9cfe1c255e855383e83372687ada48fff</a></td><td><a href="/pages/d03b47babbad2abc249a0979b2d44af1f02bbd20">/pages/d03b47babbad2abc249a0979b2d44af1f02bbd20</a></td></tr></tbody></table>

<table data-view="cards" data-full-width="false"><thead><tr><th></th><th></th><th data-hidden data-card-cover data-type="image">封面图</th><th data-hidden data-card-target data-type="content-ref"></th></tr></thead><tbody><tr><td><strong>Dynamic V3.0 GGUF</strong></td><td>面向 Qwen3.8 的全新 Dynamic 3.0 Unsloth 量化版。</td><td><a href="/files/a5aed5fcac6af837f56214ac58cfdae830e04f6b">/files/a5aed5fcac6af837f56214ac58cfdae830e04f6b</a></td><td><a href="/pages/e658f01212ed739b6cc1648a22333767661730a1">/pages/e658f01212ed739b6cc1648a22333767661730a1</a></td></tr><tr><td><strong>Meta Muse Glimmer</strong></td><td>Meta 全新的 30B 开放模型。</td><td><a href="/files/ed5a851fa9d81ba1cb05f33b85fdd6fcd68ebc71">/files/ed5a851fa9d81ba1cb05f33b85fdd6fcd68ebc71</a></td><td><a href="/pages/109c22d0c1d136d31076df8bc2667ebda88e2601">/pages/109c22d0c1d136d31076df8bc2667ebda88e2601</a></td></tr><tr><td><strong>DeepSeek-V4-Flash-0731</strong></td><td>7 月 31 日发布的新 284B 模型。</td><td><a href="/files/d4dd0c7eb8c64e6f0e6db819bc0898ec0475823e">/files/d4dd0c7eb8c64e6f0e6db819bc0898ec0475823e</a></td><td><a href="/pages/8cefdd6ae681e7680f74c7b6a344fd7c21933ee8">/pages/8cefdd6ae681e7680f74c7b6a344fd7c21933ee8</a></td></tr></tbody></table>

### ⚡ 快速开始

Unsloth 支持 MacOS、Linux、 [Windows](/docs/zh/kuai-su-kai-shi/install/windows-installation.md), [NVIDIA](/docs/zh/kuai-su-kai-shi/install/pip-install.md), [AMD](/docs/zh/kuai-su-kai-shi/install/amd.md)、Intel 以及 CPU 配置。参见： [Unsloth 要求](/docs/zh/kuai-su-kai-shi/fine-tuning-for-beginners/unsloth-requirements.md)。下载适用于你操作系统的原生桌面应用：

<a href="https://unsloth.ai/download" class="button primary" data-icon="down-to-bracket">下载 Unsloth Desktop</a>

* <i class="fa-apple">:apple:</i> [下载 macOS 版本](https://unsloth.ai/download/mac)
* <i class="fa-windows">:windows:</i> [下载 Windows 版本](https://unsloth.ai/download/windows)
* <i class="fa-linux">:linux:</i> [下载 Linux 版本](https://unsloth.ai/download/linux)

或者，如果你更喜欢手动安装：

**MacOS、Linux、WSL：**

```bash
curl -fsSL https://unsloth.ai/install.sh | sh
```

**Windows PowerShell：**

```bash
irm https://unsloth.ai/install.ps1 | iex
```

{% columns %}
{% column width="50%" %}
{% content-ref url="/pages/8d4117b244a368b8f80b1a9d079fa31360c8e823" %}
[Complete LLM Directory](/docs/zh/mo-xing/tutorials.md)
{% endcontent-ref %}

{% content-ref url="/pages/e722e86e330786e5915445b91f900d9f9e0ba067" %}
[Fine-tuning Guide](/docs/zh/kuai-su-kai-shi/fine-tuning-llms-guide.md)
{% endcontent-ref %}
{% endcolumn %}

{% column width="50%" %}
{% content-ref url="/pages/20805f6881460e7c3a088cff24acf0f1090f1984" %}
[Models](/docs/zh/kuai-su-kai-shi/unsloth-model-catalog.md)
{% endcontent-ref %}

{% content-ref url="/pages/2c2bb53a273009e389791ded9e28dd4769a55051" %}
[Unsloth API](/docs/zh/ji-chu/api.md)
{% endcontent-ref %}
{% endcolumn %}
{% endcolumns %}

然后在浏览器中打开 `http://127.0.0.1:8888` （或者你自己的特定 URL）。

### 👾 Unsloth 启动

{% columns %}
{% column width="58.333333333333336%" %}
[Unsloth 启动](/docs/zh/ji-cheng/unsloth-start.md) 让你连接 [Claude Code](/docs/zh/ji-chu/claude-code.md), [Codex](/docs/zh/ji-chu/codex.md) 以及其他代理到本地模型，通过 `unsloth start` 命令。

启动 Unsloth，加载一个模型，打开你的项目文件夹，然后运行：

```bash
unsloth start claude
```

将 `claude` 替换为下面任意一个代理：
{% endcolumn %}

{% column width="41.666666666666664%" %}

<figure><img src="/files/86e035acbaafcca2b57e4bb3f4191ad68bf30547" alt="" width="563"><figcaption><p>Claude Code 在本地运行 Qwen3.5。</p></figcaption></figure>
{% endcolumn %}
{% endcolumns %}

| 代理                                                                 | 命令                       |
| ------------------------------------------------------------------ | ------------------------ |
| <i class="fa-claude">:claude:</i> Claude Code                      | `unsloth start claude`   |
| <i class="fa-openai">:openai:</i> OpenAI Codex                     | `unsloth start codex`    |
| <i class="fa-caduceus">:caduceus:</i> Hermes Agent                 | `unsloth start hermes`   |
| <i class="fa-lobster">:lobster:</i> OpenClaw                       | `unsloth start openclaw` |
| <i class="fa-rectangle-vertical">:rectangle-vertical:</i> OpenCode | `unsloth start opencode` |

### 🦥 为什么选择 Unsloth？

* 我们直接与以下团队合作： [gpt-oss](https://docs.unsloth.ai/new/gpt-oss-how-to-run-and-fine-tune#unsloth-fixes-for-gpt-oss), [Qwen3](https://www.reddit.com/r/LocalLLaMA/comments/1kaodxu/qwen3_unsloth_dynamic_ggufs_128k_context_bug_fixes/), [Llama 4](https://github.com/ggml-org/llama.cpp/pull/12889), [Mistral](https://huggingface.co/mistralai/Mistral-Medium-3.5-128B/discussions/18), [Gemma 1-3](https://news.ycombinator.com/item?id=39671146) 和 [Phi-4](https://unsloth.ai/blog/phi4)，在此过程中我们 **修复了关键 bug** 这些修复大幅提升了模型准确率。比如 Andrej Karpathy 还 [称赞过我们的工作](https://x.com/karpathy/status/1765473722985771335).
* Unsloth 简化了本地训练、推理、数据处理和部署
* Unsloth 支持 500+ 个模型的推理和训练： [视觉](/docs/zh/ji-chu/vision-fine-tuning.md), [TTS](/docs/zh/ji-chu/text-to-speech-tts-fine-tuning.md), [嵌入](/docs/zh/ji-chu/embedding-finetuning.md), [强化学习](/docs/zh/kuai-su-kai-shi/reinforcement-learning-rl-guide.md)

### ⭐ 功能

Unsloth 让你运行和训练文本模型、 [音频](https://unsloth.ai/docs/basics/text-to-speech-tts-fine-tuning), [嵌入](https://unsloth.ai/docs/new/embedding-finetuning), [视觉](https://unsloth.ai/docs/basics/vision-fine-tuning) 等等。Unsloth 为推理和训练提供了许多关键功能：

#### 推理

* [自我修复式工具调用](/docs/zh/xin/studio/chat.md#auto-healing-tool-calling)  / 网页搜索和使用 [将 Unsloth 作为 API 使用](/docs/zh/ji-chu/api.md).
* 将你的本地模型连接到任意代理： [Claude Code](/docs/zh/ji-chu/claude-code.md), [Codex](/docs/zh/ji-chu/codex.md), [Hermes](/docs/zh/ji-cheng/hermes-agent.md) 等等。
* 搜索 + 下载 + 运行任何模型，如 GGUF、LoRA 适配器、safetensors。
* [自动推理参数](/docs/zh/xin/studio/chat.md#auto-parameter-tuning) 调优并编辑聊天模板。
* [导出或保存](/docs/zh/xin/studio/export.md) 你的模型为 GGUF、16 位 safetensor 等格式。
* [对比输出](/docs/zh/xin/studio/chat.md#model-arena) 并排比较两个不同模型的输出。

#### 训练

* 训练并 [强化学习](/docs/zh/kuai-su-kai-shi/reinforcement-learning-rl-guide.md) 让 500+ 个模型的训练速度提升约 2 倍，显存占用减少约 70%（不损失准确率）
* 支持全量微调、预训练、4 位、16 位和 FP8 训练。
* [自动创建数据集](/docs/zh/xin/studio/data-recipe.md) 从 PDF、CSV、DOCX 文件中生成。可在可视化节点工作流中编辑数据。
* 可观测性：实时监控训练，跟踪损失、GPU 使用情况，自定义图表
* 最高效的 [**强化学习**](/docs/zh/kuai-su-kai-shi/reinforcement-learning-rl-guide.md) 库，GRPO 的显存占用减少 80%， [FP8](/docs/zh/kuai-su-kai-shi/reinforcement-learning-rl-guide/fp8-reinforcement-learning.md) 等等。
* [多 GPU](/docs/zh/ji-chu/multi-gpu-training-with-unsloth.md) 目前可用，但更强大的版本即将推出！

### **最新模型**

<table data-column-title-hidden data-card-wrap="false" data-view="cards"><thead><tr><th></th><th data-hidden data-card-cover data-type="image">封面图</th><th data-hidden data-card-target data-type="content-ref"></th></tr></thead><tbody><tr><td><h4>Qwen3.8</h4></td><td><a href="/files/0fb600e9cfe1c255e855383e83372687ada48fff">/files/0fb600e9cfe1c255e855383e83372687ada48fff</a></td><td><a href="/pages/d03b47babbad2abc249a0979b2d44af1f02bbd20">/pages/d03b47babbad2abc249a0979b2d44af1f02bbd20</a></td></tr><tr><td><h4>Meta Muse Glimmer</h4></td><td><a href="/files/ed5a851fa9d81ba1cb05f33b85fdd6fcd68ebc71">/files/ed5a851fa9d81ba1cb05f33b85fdd6fcd68ebc71</a></td><td></td></tr><tr><td><h4>Kimi K3</h4></td><td><a href="/files/034abeba346d5512a9fa015c1b89656d4230f2c8">/files/034abeba346d5512a9fa015c1b89656d4230f2c8</a></td><td><a href="/pages/5b05e4653b1bff8c668dedc50b55c1adb50beb16">/pages/5b05e4653b1bff8c668dedc50b55c1adb50beb16</a></td></tr><tr><td><h4>DeepSeek V4</h4></td><td><a href="/files/67b015dc396f26780060fa85863cc7660b18228a">/files/67b015dc396f26780060fa85863cc7660b18228a</a></td><td><a href="/pages/8cefdd6ae681e7680f74c7b6a344fd7c21933ee8">/pages/8cefdd6ae681e7680f74c7b6a344fd7c21933ee8</a></td></tr><tr><td><h4>Qwen3.6</h4></td><td><a href="/files/3633b4d39ba4c75356f2106d848b79d2dbba8359">/files/3633b4d39ba4c75356f2106d848b79d2dbba8359</a></td><td><a href="/pages/213bd08e4302b621f4392f7ee38decb275ffab02">/pages/213bd08e4302b621f4392f7ee38decb275ffab02</a></td></tr><tr><td><h4>GLM-5.2</h4></td><td><a href="/files/9648d94ff49ae13dc1a6da9ac1e479f53756c7e9">/files/9648d94ff49ae13dc1a6da9ac1e479f53756c7e9</a></td><td><a href="/pages/986f178fb44971bd081c36c351922d60469ba3ee">/pages/986f178fb44971bd081c36c351922d60469ba3ee</a></td></tr><tr><td><h4>Gemma 4</h4></td><td><a href="/files/4cf8e9a56e43aacbcc50d5a8e5f62a33b061cd8d">/files/4cf8e9a56e43aacbcc50d5a8e5f62a33b061cd8d</a></td><td><a href="/pages/10f714f4a513e0d0a86b6f9d5945f9014729b035">/pages/10f714f4a513e0d0a86b6f9d5945f9014729b035</a></td></tr><tr><td><h4>MiniMax M3</h4></td><td><a href="/files/5e1a20606c6f93a6455f7dc971cd6123d39b58e7">/files/5e1a20606c6f93a6455f7dc971cd6123d39b58e7</a></td><td><a href="/pages/c93747ce9d408969d5a40287410c760083372c5c">/pages/c93747ce9d408969d5a40287410c760083372c5c</a></td></tr><tr><td><h4>DiffusionGemma</h4></td><td><a href="/files/c59173b40aad29496c4f99dad86797ecfc5dcc1c">/files/c59173b40aad29496c4f99dad86797ecfc5dcc1c</a></td><td><a href="/pages/f2f653bcccc655dc99b2d3e3b6988523765d5c44">/pages/f2f653bcccc655dc99b2d3e3b6988523765d5c44</a></td></tr></tbody></table>

### **视频演示**

{% embed url="<https://www.youtube.com/watch?v=_1hgTTuber4>" %}

{% columns %}
{% column width="50%" %}
{% content-ref url="/pages/e1e43893beb1c3e2a075324e9a00800315b2e1a3" %}
[Unsloth 更新](/docs/zh/xin/changelog.md)
{% endcontent-ref %}

{% content-ref url="/pages/9a72670992feb75def412a693565c84a88c8a266" %}
[推理与部署](/docs/zh/ji-chu/inference-and-deployment.md)
{% endcontent-ref %}
{% endcolumn %}

{% column width="50%" %}
{% content-ref url="/pages/4180d003e52f27564574c43829dd41c1236913b6" %}
[Unsloth Start](/docs/zh/ji-cheng/unsloth-start.md)
{% endcontent-ref %}

{% content-ref url="/pages/e658f01212ed739b6cc1648a22333767661730a1" %}
[Dynamic 3.0 GGUFs](/docs/zh/ji-chu/dynamic-3.0-ggufs.md)
{% endcontent-ref %}
{% endcolumn %}
{% endcolumns %}

<figure><img src="/files/e4b5cbe009022b39c4c31fc9c211fd0ea61c60fb" alt="" width="188"><figcaption></figcaption></figure>


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://unsloth.ai/docs/zh/docs.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.