🔶 openKylin：面向 AI Agent 的 Linux 桌面

一些 Linux 发行版正在回避 AI；中国的 openKylin 则积极拥抱 AI。

布拉格——许多 Linux 发行版，例如 Gentoo、Void Linux 以及 System76 的 COSMIC 桌面，都不喜欢 AI。另一方面，Linus Torvalds 在 Open Source Summit Europe 上表示：“我实际上非常喜欢使用 AI。”此外，还有由社区开发的中国 Linux 发行版 openKylin，它对此毫不讳言：其核心方向就是将 AI 置于中心位置。

正如 openKylin 执行董事刘敏在 Open Source Summit Europe 的主题演讲中所解释的那样，该操作系统正从管理计算资源，转向通过其 KylinBot AI Agent 框架组织 AI 能力，并最终组织用户意图。

借助 KylinBot，重点不再是传统的手动桌面操作体验，而是意图管理：理解用户想做什么，协调模型、数据、工具和设备，并在既定权限范围内执行由此产生的任务。

这种转变是从资源管理走向能力管理，但刘敏指出：“资源管理仍然是基础。”简而言之，最新发布的 openKylin 3.0 试图将 AI 构建进桌面的操作架构，而不只是为 Linux 发行版捆绑一个聊天机器人。

具体而言，openKylin 表示，KylinBot 可以通过语音和聊天理解用户请求，在多次交互之间保留上下文，并利用系统能力帮助用户完成工作。通过集成自动化工具 Kylin-CUA，KylinBot 可以访问系统设置和设备管理等能力，使 AI 不再局限于回答问题，而是能够执行任务。

在实际使用中，KylinBot 提供 Agent 框架，而 Xiao K 则在其之上提供助手界面。可用的系统技能涵盖输入设备、显示设置、网络、蓝牙外设、电源管理及其他桌面控制功能。例如，用户可以通过与 KylinBot 对话或聊天，请求切换网络连接、关闭蓝牙或更改键盘布局，而无需操作常规设置界面。

该项目还集成了 Kylin-CUA。这一计算机使用自动化工具将 Agent 的访问范围扩展至系统设置和设备管理功能。它还支持第三方 Agent 工具，例如 OpenClaw 和 WorkBuddy。

这一集成是最新版本 openKylin 3.0 的核心。在该版本中，Linux 桌面成为 Agent 可以操作的对象，而不再只是运行 Agent 聊天窗口的场所。

刘敏在主题演讲中以分层架构阐述了这些功能。熟悉的 Linux、驱动程序和硬件资源仍处于基础层。其上是本地推理能力与共享 AI 运行时服务，随后是负责规划任务和协调工具的 Agent。计算机使用自动化则进一步扩展了这些 Agent 对传统图形应用程序的操作范围。

在底层方面，除 AI 工作外，openKylin 也包含常规的 Linux 平台升级。例如，openKylin 3.0 使用 Linux 7.0，优化了 180 多个核心组件，改进了对 RISC-V 的支持，并在 Linux 系统工具中扩大 Rust 的应用。

刘敏补充称，该项目包含可按内容和描述定位信息的智能搜索功能。这降低了用户对文件具体位置的依赖；智能工作空间则围绕一项任务收集相关文件。这些功能被作为该项目探索 AI 如何改变桌面交互的一部分提出，而非作为已被独立验证的成果。

OpenKylin 仍在持续开发中。刘敏解释道：“开源让 AI 能力变得可获取，但可获取只是开始。”她补充说，要让它真正有用，还需要系统工程：理解模型如何运行、如何连接数据和应用程序，以及如何在不同硬件环境中工作。

这一中国开源 Linux 发行版主要面向桌面计算，同时也是商业版麒麟操作系统的技术上游。

OpenKylin 也在探索传统桌面以外的应用场景，包括智能设备、工业自动化、机器人、高性能计算和云环境。这些仍属于发展方向；它尚未为每一种工作负载提供同等成熟的平台。

其开发者所认识到的另一项问题是安全性。毕竟，一个能够修改设置和操作应用程序的 AI Agent 会带来诸多担忧。刘敏将 Agent 安全与治理列为持续优先事项，其中包括由谁授权操作、允许哪些操作，以及人类如何进行干预。

正如 openKylin 技术委员会主席吴庆波所补充的那样，openKylin 正在利用一种根本性的架构变化。“从‘AI 作为应用’转向‘AI 作为操作系统’，并不只是功能升级，而是底层架构的变革。”展望未来，他预计其开放生态和架构将使开发者能够“以更透明、更安全的方式”扩展智能能力，同时适配更多硬件，例如 RISC-V。

或者，正如刘敏在结束语中所说：“开源汇聚创新。操作系统让智能变得可用。”对于 openKylin 3.0 而言，问题在于：一个开放的 Linux 桌面能否将 AI 从独立应用转变为协调一致的系统能力，并同时提供实用的性能和有意义的人类控制。

Some Linux distros are steering clear of AI. China’s openKylin embraces AI.

Prague — Many Linux distros such as Gentoo, Void Linux, and System76’s COSMIC desktop hate AI. Linus Torvalds, on the other hand, said at Open Source Summit Europe, “I actually really like using AI.” Then, there’s China’s openKylin community-developed Linux distribution, which makes no bones about it. They’re all about putting AI front and center.

As openKylin executive director Liu Min explained in her keynote at Open Source Summit Europe, the operating system is moving beyond managing computing resources to organizing AI capabilities and, ultimately, users’ intent through its KylinBot AI agent framework.

With KylinBot, instead of the usual hands-on desktop experience, it’s all about intent management: understanding what a user wants, coordinating models, data, tools, and devices, and executing the resulting tasks within defined permissions.

The shift is from resource management to capability management, but Liu observed: “Resource management remains the foundation.” In short, openKylin 3.0, its latest release, attempts to build AI into the desktop’s operating architecture rather than merely bundling a chatbot with a Linux distribution.

Specifically, openKylin says KylinBot can understand user requests by voice and chat, retain context across interactions, and use system capabilities to help you do work. By integrating Kylin-CUA, an automation tool, KylinBot can access capabilities such as system settings and device management, moving AI beyond answering questions toward executing tasks.

What this looks like in practice is that KylinBot provides the agent framework, while Xiao K supplies an assistant interface built on top of it. Available system skills cover input devices, display settings, networking, Bluetooth peripherals, power management, and other desktop controls. For example, you can request a different network connection, disable Bluetooth, or change a keyboard layout by talking or chatting with KylinBot, without touching the usual settings controls.

The project also integrates Kylin-CUA. This computer-use automation tool extends agent access to system settings and device-management functions. It also supports third-party agent tools, such as OpenClaw and WorkBuddy.

This integration is central to openKylin 3.0, the latest release. In it, the Linux desktop becomes something an agent can operate, not simply somewhere an agent’s chat window runs.

Liu’s keynote placed these features in a layered architecture. Familiar Linux, drivers, and hardware resources remain the foundation. Above them sit local inference capabilities and shared AI runtime services, followed by agents that plan tasks and coordinate tools. Computer-use automation extends those agents’ reach to traditional graphical applications.

Under the hood, openKylin includes conventional Linux-platform upgrades alongside its AI work. For example, openKylin 3.0 uses Linux 7.0, optimizes more than 180 core components, improves RISC-V support, and expands Rust’s use in its Linux system tools.

Liu added that the project includes intelligent search that locates information by content and description. This reduces reliance on knowing a file’s location and on intelligent workspaces by collecting relevant files around a task. These were presented as part of the project’s exploration of how AI changes desktop interaction, rather than as independently demonstrated results.

OpenKylin is still a work in progress. Liu explained, “Open source makes AI capability accessible, but accessibility is only the beginning,” Making it useful, she added, requires system engineering: understanding how models run, connect to data and applications, and work across different hardware environments.

This Chinese open-source Linux distribution primarily focuses on desktop computing. It serves as a technical upstream for the commercial Kylin operating system.

OpenKylin is also exploring uses beyond the traditional desktop. These include smart devices, industrial automation, robotics, high-performance computing, and cloud environments. Those are development directions. It doesn’t yet provide an equally mature platform for every workload.

Another concern its developers are aware of is its security. After all, an AI agent that can change settings and operate applications raises numerous concerns. Liu identified agent security and governance as a continuing priority, including who authorizes actions, which actions are permitted, and how a human can intervene.

As openKylin Technical Committee chair Wu Qingbo added, openKylin is exploiting a fundamental architectural change. “The shift from ‘AI as an application’ to ‘AI as an operating system’ is not simply a feature upgrade, but a transformation of the underlying architecture.” Looking ahead, he expects its open ecosystem and architecture will enable developers to extend intelligent capabilities “more transparently and securely” while accommodating more hardware, such as RISC-V.

Or, as Liu’s closing statement put it: “Open source brings innovation together. Operating systems make intelligence usable.” For openKylin 3.0, the question is whether an open Linux desktop can transform AI from a separate application into a coordinated system capability—and do so with useful performance and meaningful human control.