diff --git a/docs/rock4/rock4d/application-dev/npu/kiln.md b/docs/rock4/rock4d/application-dev/npu/kiln.md
new file mode 100644
index 000000000..621c0fbfb
--- /dev/null
+++ b/docs/rock4/rock4d/application-dev/npu/kiln.md
@@ -0,0 +1,223 @@
+---
+sidebar_position: 25
+
+doc_kind: page
+source_of_truth: local
+title: "Kiln：在主线内核上运行本地 LLM + 视觉（NPU）"
+description: "使用 Kiln 在 ROCK 4D 上，于主线 Linux 内核运行厂商 RKLLM/RKNN NPU 栈：一条命令装好本地大模型对话、图像识别，以及 OpenAI 兼容 API（可接 Open WebUI）。"
+locale: zh
+board: rock4d
+task_type: getting-started
+last_verified: 2026-07-11
+---
+
+# Kiln：在主线内核上运行本地 LLM + 视觉（NPU）
+
+本教程带你在 **ROCK 4D（RK3576）** 上，用 [Kiln](https://github.com/gahingwoo/kiln) 从零搭起一个
+**私有、离线**的 AI 环境：在 NPU 上跟本地大模型对话、给图片做分类 / 目标检测，并暴露成 **OpenAI
+兼容的 API**（可直接接 Open WebUI 这类 ChatGPT 式网页）。整个过程数据不出设备、无需联网推理、不用
+API key。
+
+Kiln 的特别之处是它运行在**主线 Linux 内核**（`linux-7.1.3`）上，而非厂商 6.1 BSP：它把厂商 GPL
+`rknpu` 驱动以 out-of-tree 方式编译，配上闭源的 `librkllmrt`（大模型）/ `librknnrt`（视觉）运行时，
+再加一小组针对 RK3576 NPU 的内核补丁（时钟 / 电源域 / 双 IOMMU），让 NPU 在主线内核上正常工作。
+
+装好后你会得到这些命令：
+
+| 命令           | 作用                                                |
+| -------------- | --------------------------------------------------- |
+| `kiln`         | 总入口菜单：对话 / 视觉 / 模型 / 服务 / 设置 / 诊断 |
+| `kiln-chat`    | 在 NPU 上跟本地大模型对话                           |
+| `kiln-vision`  | 图像分类 / YOLO 目标检测                            |
+| `kiln-serve`   | OpenAI 兼容 HTTP API（接 Open WebUI、LangChain 等） |
+| `kiln-convert` | 在板子上把 ONNX 转成 `.rknn`                        |
+| `kiln-config`  | 配置 TUI（模型 / 参数 / 服务）                      |
+| `kiln-doctor`  | 健康检查                                            |
+
+:::note
+Kiln 是独立的第三方开源项目，不属于 Radxa 官方支持范围。问题与反馈请到其
+[仓库](https://github.com/gahingwoo/kiln)。目前它只在 ROCK 4D（RK3576）上做过完整验证。
+:::
+
+## Prerequisites
+
+- 一块 **ROCK 4D（RK3576）**，运行 **Armbian**。
+- 可用网络（首次安装需要联网下载内核与运行时；**以太网最稳**）。
+- 模型自备（Kiln 不附带任何模型）：
+  - **视觉**：无需自备，下面用 `kiln-convert` 在板上现转一个 MobileNet 分类器。
+  - **大模型**：准备一个与 `librkllmrt` **1.2.0** 匹配的 `*-rk3576-w4a16.rkllm`（例如 Qwen2.5-1.5B
+    或 Llama-3.2-1B），稍后放进 `/opt/models`。
+
+## Steps
+
+### 1. 安装
+
+在 ROCK 4D 上执行这条命令：
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/gahingwoo/kiln/main/scripts/kiln-install.sh | bash
+```
+
+它是**免手动**的：预下载所需内容、装上 Kiln 主线内核，然后**自己重启两次（共约 10–15 分钟）**完成
+安装并进入就绪系统。
+
+:::caution
+安装过程中板子会自动重启两次，**请勿断电**。两次重启之间板载 Wi-Fi 会短暂断开——这是预期的，安装的
+第二阶段离线完成、不需要网络。
+:::
+
+:::tip
+想自己掌控重启，用**手动模式**——把脚本管道给带环境变量的 bash（它会告诉你何时重启、何时再跑一遍）：
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/gahingwoo/kiln/main/scripts/kiln-install.sh | KILN_MANUAL=1 bash
+```
+
+:::
+
+### 2. 验证安装
+
+装完登录后，跑健康检查：
+
+```bash
+kiln-doctor
+```
+
+关键项应为 `[ OK ]`：
+
+```text
+NPU driver (rknpu)
+  [ OK ] rknpu loaded
+  [ OK ] render node present (/dev/dri/renderD128)
+
+NPU MMU state (dmesg)
+  [ OK ] all four MMU banks enabled (st=0x19/0x19/0x19/0x19)
+
+Runtimes
+  [ OK ] librkllmrt.so  RKLLM SDK (version: 1.2.0 ...)
+  [ OK ] librknnrt.so   librknnrt version: 2.3.0 ...
+```
+
+看到 `rknpu loaded`、`renderD128`、四个 MMU bank `0x19`、以及两个运行时版本，就说明 NPU 已就绪。
+
+### 3. 图像识别（视觉）
+
+先在板上现转一个 MobileNet 分类器（首次会安装 rknn-toolkit2，约几分钟），再跑一张测试图：
+
+```bash
+kiln-convert mobilenet --set-active
+kiln-vision /opt/models/test.jpg
+```
+
+预期输出（NPU 推理约 6 ms）：
+
+```text
+top-5 of 1000 classes  (NPU inference 5.9 ms):
+  1. [ 494] chime, bell, gong            18.6719
+  ...
+[bench] rknn inference: 5.9 ms (169.5 fps)
+```
+
+想做**目标检测**（YOLO），转一个检测器并给第二个路径参数保存带框图：
+
+```bash
+kiln-convert yolov8n --set-active
+kiln-vision /opt/models/test.jpg out.jpg
+```
+
+### 4. 大模型对话
+
+把一个匹配 `librkllmrt` 1.2.0 的 `*-rk3576-w4a16.rkllm` 放进 `/opt/models`（从你的电脑 `scp` 过去），
+然后：
+
+```bash
+kiln-chat
+```
+
+`kiln-chat` 会自动发现 `/opt/models` 里的 `.rkllm`，每轮回复后打印一行速度基准：
+
+```text
+you > 用一句话介绍 RK3576。
+bot > ...
+[bench] tokens=42  prefill(TTFT)=180 ms  decode=12.8 tok/s  total=...
+```
+
+:::tip
+在提示符输入单独一个 `/` 回车，会弹出所有斜杠命令的菜单（`/model` 换模型、`/system` 设系统提示等），
+不用记也不用先打 `/help`。
+:::
+
+### 5. 接入 Open WebUI / OpenAI 客户端
+
+`kiln-serve` 把大模型放到一个 OpenAI 兼容的 HTTP API 后面，任何 OpenAI 客户端换个 base URL 就能用。
+
+在板子上启动服务：
+
+```bash
+kiln-serve
+```
+
+它会打印**填好板子 IP 的连接串**，直接抄用即可：
+
+```text
+kiln-serve: ready [chat+classify]. Listening on 0.0.0.0:8080  (OpenAI /v1)
+  -> Open WebUI / OpenAI:  OPENAI_API_BASE_URL=http://<board-ip>:8080/v1   (API key: any)
+  -> test:                 curl http://<board-ip>:8080/v1/models
+```
+
+在**你的电脑**上用 Docker 跑 [Open WebUI](https://github.com/open-webui/open-webui)，指向板子（把
+IP 换成上面打印的那个）：
+
+```bash
+docker run -d -p 3000:8080 -e OPENAI_API_BASE_URL=http://<board-ip>:8080/v1 -e OPENAI_API_KEY=kiln ghcr.io/open-webui/open-webui:main
+```
+
+然后打开 `localhost:3000`，模型选择器里就有板子上的模型，对话逐 token 从 NPU 流式输出。
+
+用 `openai` SDK 也一样（只改 base URL）：
+
+```python
+from openai import OpenAI
+client = OpenAI(base_url="http://<board-ip>:8080/v1", api_key="kiln")
+r = client.chat.completions.create(model="kiln", messages=[{"role": "user", "content": "hello"}])
+print(r.choices[0].message.content)
+```
+
+:::tip
+`[server] host` 必须是 `0.0.0.0` 才能从别的机器连；若是 `127.0.0.1` 只有本机能连（`kiln-doctor`
+会提醒）。改法：`kiln-config` → Server。
+:::
+
+### 6. 转换你自己的模型
+
+`kiln-convert` 在板子上把 ONNX 转成版本匹配的 `.rknn`，无需 x86 机器：
+
+```bash
+kiln-convert mobilenet            # 拉 MobileNet + 转（分类）
+kiln-convert yolov8n              # 拉 YOLOv8n + 转（检测；Ultralytics 为 AGPL，会先询问）
+kiln-convert ./my_model.onnx      # 转你自己的 ONNX
+```
+
+它首次会建一个私有 `rknn-toolkit2` 环境并**锁到与运行时匹配的版本**，避免版本不匹配导致加载崩溃。
+加 `--set-active` 会把结果写进配置、立即可用。
+
+## Validation
+
+到这里你应当已经能：`kiln-doctor` 全绿、`kiln-vision` 正确分类测试图（约 6 ms）、放入 `.rkllm` 后
+`kiln-chat` 能对话、`kiln-serve` + Open WebUI 能在网页里聊天。也可以直接跑 `kiln` 打开菜单逐项体验。
+
+## Troubleshooting
+
+- **先跑 `kiln-doctor`**，把完整输出贴到 issue —— 这是最有用的信息。
+- `std::out_of_range in rknn_inputs_set`：`.rknn` 用了不匹配的 rknn-toolkit2 版本转的；用
+  `kiln-convert` 重转（它会自动锁版本）。
+- 检测（YOLO）时崩溃：导出要**关 NMS**（`nms=False`）；`kiln-convert yolov8n` 会自动处理。
+- 从别的机器连不上 `kiln-serve`：把 `[server] host` 设成 `0.0.0.0`（`kiln-config` → Server）。
+- 更多报错对照 Kiln 的
+  [故障排查文档](https://github.com/gahingwoo/kiln/blob/main/docs/zh/TROUBLESHOOTING.md)。
+
+## 参考
+
+- Kiln 仓库：[gahingwoo/kiln](https://github.com/gahingwoo/kiln)
+- Kiln 中文文档：[README.zh-CN.md](https://github.com/gahingwoo/kiln/blob/main/README.zh-CN.md)
+- Open WebUI 接入：[docs/zh/OPENWEBUI.md](https://github.com/gahingwoo/kiln/blob/main/docs/zh/OPENWEBUI.md)
diff --git a/i18n/en/docusaurus-plugin-content-docs/current/rock4/rock4d/application-dev/npu/kiln.md b/i18n/en/docusaurus-plugin-content-docs/current/rock4/rock4d/application-dev/npu/kiln.md
new file mode 100644
index 000000000..ff715129b
--- /dev/null
+++ b/i18n/en/docusaurus-plugin-content-docs/current/rock4/rock4d/application-dev/npu/kiln.md
@@ -0,0 +1,243 @@
+---
+sidebar_position: 25
+
+doc_kind: page
+source_of_truth: local
+title: "Kiln: local LLM + vision on the NPU (mainline kernel)"
+description: "Use Kiln on the ROCK 4D to run the vendor RKLLM/RKNN NPU stack on a mainline Linux kernel: one command for a local LLM chat, image recognition, and an OpenAI-compatible API you can point Open WebUI at."
+locale: en
+board: rock4d
+task_type: getting-started
+last_verified: 2026-07-11
+---
+
+# Kiln: local LLM + vision on the NPU (mainline kernel)
+
+This tutorial walks you through building a **private, offline** AI setup on the
+**ROCK 4D (RK3576)** with [Kiln](https://github.com/gahingwoo/kiln): chat with a local
+LLM on the NPU, classify / detect objects in images, and expose an **OpenAI-compatible
+API** (which you can point a ChatGPT-style web UI like Open WebUI at). Nothing leaves the
+device — no cloud, no API keys.
+
+What makes Kiln different is that it runs on a **mainline Linux kernel** (`linux-7.1.3`)
+instead of the vendor 6.1 BSP: it builds the vendor GPL `rknpu` driver out-of-tree, drives
+it with the closed `librkllmrt` (LLM) / `librknnrt` (vision) runtimes, and adds a small set
+of RK3576 NPU kernel patches (clock / power-domain / two-IOMMU) so the NPU works on a
+mainline kernel.
+
+After installing you get these commands:
+
+| Command        | What it does                                                           |
+| -------------- | ---------------------------------------------------------------------- |
+| `kiln`         | Umbrella menu: chat / vision / models / serve / settings / diagnostics |
+| `kiln-chat`    | Chat with a local LLM on the NPU                                       |
+| `kiln-vision`  | Image classification / YOLO object detection                           |
+| `kiln-serve`   | OpenAI-compatible HTTP API (Open WebUI, LangChain, …)                  |
+| `kiln-convert` | Convert an ONNX to `.rknn` on the board                                |
+| `kiln-config`  | Config TUI (models / parameters / server)                              |
+| `kiln-doctor`  | Health check                                                           |
+
+:::note
+Kiln is an independent third-party open-source project, not part of official Radxa
+support. Report issues on its [repository](https://github.com/gahingwoo/kiln). It has been
+fully validated only on the ROCK 4D (RK3576).
+:::
+
+## Prerequisites
+
+- A **ROCK 4D (RK3576)** running **Armbian**.
+- A working network (the first install downloads the kernel and runtimes; **Ethernet is
+  most reliable**).
+- You supply the models (Kiln ships none):
+  - **Vision**: nothing to prepare — below you convert a MobileNet classifier on the board
+    with `kiln-convert`.
+  - **LLM**: a `*-rk3576-w4a16.rkllm` matched to `librkllmrt` **1.2.0** (e.g. Qwen2.5-1.5B
+    or Llama-3.2-1B), placed in `/opt/models` later.
+
+## Steps
+
+### 1. Install
+
+Run this on the ROCK 4D:
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/gahingwoo/kiln/main/scripts/kiln-install.sh | bash
+```
+
+It is hands-off: it pre-downloads what it needs, installs the Kiln mainline kernel, then
+**reboots itself twice (~10–15 minutes total)** to finish setup and land in a ready system.
+
+:::caution
+The board reboots itself twice during install — **do not cut power**. Onboard Wi-Fi is
+briefly down between the reboots; that is expected, as phase 2 finishes offline and does
+not need the network.
+:::
+
+:::tip
+To drive the reboots yourself, use **manual mode** — pipe the script to bash with the
+environment variable set (it tells you when to reboot and re-run):
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/gahingwoo/kiln/main/scripts/kiln-install.sh | KILN_MANUAL=1 bash
+```
+
+:::
+
+### 2. Verify the install
+
+After logging in, run the health check:
+
+```bash
+kiln-doctor
+```
+
+The key items should read `[ OK ]`:
+
+```text
+NPU driver (rknpu)
+  [ OK ] rknpu loaded
+  [ OK ] render node present (/dev/dri/renderD128)
+
+NPU MMU state (dmesg)
+  [ OK ] all four MMU banks enabled (st=0x19/0x19/0x19/0x19)
+
+Runtimes
+  [ OK ] librkllmrt.so  RKLLM SDK (version: 1.2.0 ...)
+  [ OK ] librknnrt.so   librknnrt version: 2.3.0 ...
+```
+
+`rknpu loaded`, `renderD128`, the four MMU banks at `0x19`, and both runtime versions mean
+the NPU is ready.
+
+### 3. Image recognition (vision)
+
+Convert a MobileNet classifier on the board (the first conversion installs rknn-toolkit2,
+a few minutes), then run a test image:
+
+```bash
+kiln-convert mobilenet --set-active
+kiln-vision /opt/models/test.jpg
+```
+
+Expected output (NPU inference ~6 ms):
+
+```text
+top-5 of 1000 classes  (NPU inference 5.9 ms):
+  1. [ 494] chime, bell, gong            18.6719
+  ...
+[bench] rknn inference: 5.9 ms (169.5 fps)
+```
+
+For **object detection** (YOLO), convert a detector and pass a second path to save an
+annotated image:
+
+```bash
+kiln-convert yolov8n --set-active
+kiln-vision /opt/models/test.jpg out.jpg
+```
+
+### 4. LLM chat
+
+Place a `*-rk3576-w4a16.rkllm` matched to `librkllmrt` 1.2.0 in `/opt/models` (scp it from
+your computer), then:
+
+```bash
+kiln-chat
+```
+
+`kiln-chat` auto-discovers any `.rkllm` in `/opt/models` and prints a speed benchmark after
+each reply:
+
+```text
+you > Describe the RK3576 in one sentence.
+bot > ...
+[bench] tokens=42  prefill(TTFT)=180 ms  decode=12.8 tok/s  total=...
+```
+
+:::tip
+Type a lone `/` at the prompt to open a menu of the slash commands (`/model` to switch
+model, `/system` to set a system prompt, etc.) — no need to memorize them or run `/help`.
+:::
+
+### 5. Connect Open WebUI / an OpenAI client
+
+`kiln-serve` puts the LLM behind an OpenAI-compatible HTTP API, so any OpenAI client works
+by just changing the base URL.
+
+Start the server on the board:
+
+```bash
+kiln-serve
+```
+
+It prints a **connection string with the board's IP already filled in** — copy it:
+
+```text
+kiln-serve: ready [chat+classify]. Listening on 0.0.0.0:8080  (OpenAI /v1)
+  -> Open WebUI / OpenAI:  OPENAI_API_BASE_URL=http://<board-ip>:8080/v1   (API key: any)
+  -> test:                 curl http://<board-ip>:8080/v1/models
+```
+
+On **your computer**, run [Open WebUI](https://github.com/open-webui/open-webui) with
+Docker, pointing it at the board (replace the IP with the one printed above):
+
+```bash
+docker run -d -p 3000:8080 -e OPENAI_API_BASE_URL=http://<board-ip>:8080/v1 -e OPENAI_API_KEY=kiln ghcr.io/open-webui/open-webui:main
+```
+
+Open `localhost:3000`; the board's model appears in the picker and chat streams
+token-by-token off the NPU.
+
+The `openai` SDK works the same way (just the base URL changes):
+
+```python
+from openai import OpenAI
+client = OpenAI(base_url="http://<board-ip>:8080/v1", api_key="kiln")
+r = client.chat.completions.create(model="kiln", messages=[{"role": "user", "content": "hello"}])
+print(r.choices[0].message.content)
+```
+
+:::tip
+`[server] host` must be `0.0.0.0` to accept connections from other machines; `127.0.0.1`
+is local-only (`kiln-doctor` warns about it). Change it via `kiln-config` → Server.
+:::
+
+### 6. Convert your own models
+
+`kiln-convert` turns an ONNX into a version-matched `.rknn` on the board — no x86 machine:
+
+```bash
+kiln-convert mobilenet            # pull MobileNet + convert (classify)
+kiln-convert yolov8n              # pull YOLOv8n + convert (detect; Ultralytics is AGPL, it asks)
+kiln-convert ./my_model.onnx      # convert your own ONNX
+```
+
+On first use it builds a private `rknn-toolkit2` environment **pinned to your runtime
+version**, so a converted model can't be version-mismatched. Add `--set-active` to write
+the result into the config for immediate use.
+
+## Validation
+
+At this point you should be able to: get an all-green `kiln-doctor`, classify the test image
+with `kiln-vision` (~6 ms), chat with `kiln-chat` once a `.rkllm` is in place, and chat from
+a web page via `kiln-serve` + Open WebUI. You can also just run `kiln` for a menu of all of
+the above.
+
+## Troubleshooting
+
+- **Run `kiln-doctor` first** and paste its full output into an issue — it's the most useful
+  thing to include.
+- `std::out_of_range in rknn_inputs_set`: the `.rknn` was converted with a mismatched
+  rknn-toolkit2 version; reconvert with `kiln-convert` (it pins the matching version).
+- Crash on a YOLO model: export with **NMS off** (`nms=False`); `kiln-convert yolov8n`
+  handles this for you.
+- Can't reach `kiln-serve` from another machine: set `[server] host` to `0.0.0.0`
+  (`kiln-config` → Server).
+- For more errors see Kiln's
+  [troubleshooting guide](https://github.com/gahingwoo/kiln/blob/main/docs/TROUBLESHOOTING.md).
+
+## References
+
+- Kiln repository: [gahingwoo/kiln](https://github.com/gahingwoo/kiln)
+- Kiln documentation: [README.md](https://github.com/gahingwoo/kiln/blob/main/README.md)
+- Open WebUI integration: [docs/OPENWEBUI.md](https://github.com/gahingwoo/kiln/blob/main/docs/OPENWEBUI.md)