openKylin Pitched Linux Rebuilt Around AI Agents at OSS Europe 2026

On October 8 at Open Source Summit Europe 2026 in Prague, openKylin Executive Director Min Liu used a main-stage keynote to introduce KART, a unified AI runtime, and a system-level agent execution environment — the community's most concrete bid yet to reposition Linux as an operating system built around AI rather than merely running it.

The keynote, titled "From AI Runtime to System Agents: Reimagining Linux for the AI Era with openKylin," framed a specific argument: the bottleneck in AI computing has moved from model capability to system infrastructure. Jianfeng Li, a member of the community's Technical Committee, reinforced the hardware side with a lightning talk titled "Building an AI-Ready Linux Desktop on RISC-V."

openKylin is an open-source operating system community originating in China. By its own figures as of October 2026, it counts approximately 27 million deployments, 2.9 million active users, and more than 29,000 contributors worldwide. Those numbers are self-reported and unaudited, a caveat worth holding onto as the project pitches itself to a global audience.

Why does the OS layer need to change at all?

Liu's diagnosis rests on a structural gap. Open frameworks such as PyTorch, vLLM, and llama.cpp supply reusable engineering components, and open-weight models — including those from DeepSeek — lower the barrier to model adaptation. Yet model capabilities remain disconnected from desktop tasks, AI stacks are fragmented, and hardware–software integration stays highly complex.

According to openKylin, that fragmentation is not a defect of any single vendor. The community frames it as a shared challenge for every operating system transitioning into the AI era, not one specific to a particular company or project.

The stakes are straightforward. If AI workloads increasingly arrive as autonomous agents rather than applications, the OS is the layer that must schedule, sandbox, and audit them. Whoever defines that substrate shapes how third-party AI integrates with desktop computing.

What is 'intent management'?

openKylin's answer is a paradigm shift it calls moving from "resource management" to "intent management." Traditional kernels manage CPUs, memory, and files. An AI-era operating system, in the project's formulation, must understand user goals and decompose them into coordinated, observable system actions — a closed loop of understanding, planning, execution, and observation.

That is an ambitious claim, and the gap between slogan and shipped code is where skepticism belongs. What openKylin demonstrated at the summit is an architecture of three layers:

• Runtime layer: KART, the unified AI runtime, plus the kytensor local inference engine (software that runs AI models directly on the device), providing shared inference services.
• Agent execution layer: an agent execution environment supporting KylinBot, Kylin-CUA (a computer-use agent that operates desktop interfaces), and third-party agents.
• Capability layer: API tools and desktop GUI skills that expose existing Linux capabilities, governed by permission boundaries and execution monitoring.

The permission and monitoring language matters more than the marketing. Agent systems that can act on the desktop raise real security questions, and openKylin says controllability and observability of agent actions are among its three stated investment priorities going forward. The proof will be in the implementation details, which the keynote did not fully specify.

In plain words: openKylin wants the OS itself — not any single app — to be the place where AI models, agents, and system tools meet, with rules attached.

What shipped in openKylin 3.0?

The architecture is not purely theoretical. In openKylin 3.0, the project has integrated three desktop features that operationalize the intent-management pitch:

• AI Assistant: interaction shifts from "finding features" to "expressing intent."
• Global Intelligent Search: search moves from location-based lookup to semantic discovery (finding content by meaning rather than by file path).
• Intelligent Space: information is organized around tasks rather than folders.

AI PCs and other devices running openKylin 3.0 were on display at the community's summit booth, where attendees could try the system directly. Demonstrations at a conference are not the same as sustained daily-driver use, but the features exist in shipping software rather than slideware.

How far does the RISC-V work go?

The most quantifiable claims in Prague concerned open hardware. Focusing on hardware enablement, RISC-V Vector (RVV) extension optimization, AI runtimes, and heterogeneous computing, openKylin reported:

• Inference acceleration: RVV kernel optimizations deliver a 2–5× speedup in FP32 inference with ONNX Runtime.
• Large-model support: the AI stack supports CPU inference with vLLM on SpacemiT K3 hardware.
• Multiple distribution formats: delivery via .deb packages, Python wheels, Conda, and Docker.

The speedup figures are openKylin's own, and the source material does not disclose the baseline hardware configurations or benchmark methodology behind them. The community says the work advances through open contributions, with plans to push both RVV kernel optimizations and inference framework enablement upstream. Upstream contribution, if completed, would move these optimizations beyond openKylin and into the broader Linux and RISC-V ecosystems — the company's own framing of why the work matters beyond its user base.

In plain words: the measurable wins are on the hardware optimization side, not yet on the intent-management side.

Who else showed up in Prague?

The summit booth anchored a small ecosystem display. Haihe Laboratory presented an interactive VR experience. SpacemiT exhibited a K3 laptop with openKylin 3.0 preinstalled — a concrete data point that the RISC-V software stack runs on commercial hardware. Singapore-based Nanyang Singtech showcased its SingNova series of AI edge computing products.

The ecosystem footprint extends past the summit floor. openKylin reports more than one million lines of core code contributed to the Linux and RISC-V ecosystems, and its UKUI desktop environment ships in the official repositories of 12 Linux distributions, including Debian, Ubuntu, and Fedora. The community also runs international user groups across more than 30 countries.

What comes next?

Looking ahead, openKylin has committed to three investment areas: strengthening the runtime foundation to make AI easier to build; enhancing security and control so agent actions remain controllable and observable; and fostering an open ecosystem for more efficient collaboration among hardware, models, and applications.

Those are the right problems to name. Whether the community delivers measurable progress on all three — particularly the security and observability piece, which agent skeptics will scrutinize hardest — will determine whether "reimagining Linux for the AI era" becomes an architecture others adopt or a keynote others forget.

By James Calloway, AI In Context