diff --git a/components/nn/include/maix_nn_yolo26_depth.hpp b/components/nn/include/maix_nn_yolo26_depth.hpp
new file mode 100644
index 000000000..ac95eb438
--- /dev/null
+++ b/components/nn/include/maix_nn_yolo26_depth.hpp
@@ -0,0 +1,336 @@
+/**
+ * @author maixpy_skill
+ * @copyright Sipeed Ltd 2023-
+ * @license Apache 2.0
+ * @update 2026.8.2: Add yolo26-depth support
+ */
+
+#pragma once
+#include <cfloat>
+#include <algorithm>
+#include <array>
+#include <cmath>
+#include <cstdint>
+#include <vector>
+#include "maix_basic.hpp"
+#include "maix_nn.hpp"
+#include "maix_image.hpp"
+#include "maix_image_cmap.hpp"
+#include "maix_time.hpp"
+
+namespace maix::nn
+{
+    /**
+     * YOLO26-depth
+     * Monocular depth estimation model (yolo26n-depth etc.) inference wrapper.
+     *
+     * Usage:
+     *     nn::Yolo26Depth model("/tmp/yolo26n-depth.mud");
+     *     image::Image *img = cam.read();
+     *     image::Image *heatmap = model.get_depth_image(*img, image::FIT_CONTAIN, image::CMap::JET);
+     *     disp.show(*heatmap);
+     *     delete heatmap;
+     *
+     * Note: the model MUD file's [extra] section should have model_type=yolo26_depth, input_type=rgb.
+     * @maixpy maix.nn.Yolo26Depth
+     */
+    class Yolo26Depth
+    {
+    public:
+        /**
+         * Construct a new Yolo26Depth object
+         * @param model MUD model path, if empty, will not load model, you can call load() later.
+         *                  if not empty, will load model and will raise err::Exception if load failed.
+         * @param[in] dual_buff prepare dual input output buffer to accelarate forward, that is, when NPU is forwarding we not wait and prepare the next input buff.
+         *                      If you want to ensure every time forward output the input's result, set this arg to false please.
+         *                      Default true to ensure speed.
+         * @maixpy maix.nn.Yolo26Depth.__init__
+         */
+        Yolo26Depth(const string &model = "", bool dual_buff = true)
+        {
+            _model = nullptr;
+            _dual_buff = dual_buff;
+            _input_img_fmt = image::Format::FMT_RGB888;
+            _input_w = _input_h = 0;
+            _output_w = _output_h = 0;
+            _cmap = image::CMap::JET;
+            if (!model.empty())
+            {
+                err::Err e = load(model);
+                if (e != err::ERR_NONE)
+                {
+                    throw err::Exception(e, "load model failed");
+                }
+            }
+        }
+
+        ~Yolo26Depth()
+        {
+            if (_model)
+            {
+                delete _model;
+                _model = nullptr;
+            }
+        }
+
+        /**
+         * Load model from file, model format is .mud,
+         * MUD file should contain [extra] section, have key-values:
+         * - model_type: yolo26_depth
+         * - input_type: rgb or bgr
+         * @param model MUD model path
+         * @return error code, if load failed, return error code
+         * @maixpy maix.nn.Yolo26Depth.load
+         */
+        err::Err load(const string &model)
+        {
+            if (_model)
+            {
+                delete _model;
+                _model = nullptr;
+            }
+            _model = new nn::NN(model, _dual_buff);
+            if (!_model)
+            {
+                return err::ERR_NO_MEM;
+            }
+            auto inputs = _model->inputs_info();
+            if (inputs.empty())
+            {
+                return err::ERR_ARGS;
+            }
+            // NHWC: [1, H, W, 3];  NCHW: [1, 3, H, W]
+            if (inputs[0].shape[3] <= 4)   // NHWC
+            {
+                _input_h = inputs[0].shape[1];
+                _input_w = inputs[0].shape[2];
+            }
+            else
+            {
+                _input_h = inputs[0].shape[2];
+                _input_w = inputs[0].shape[3];
+            }
+            // input_type decides input image format
+            auto extra = _model->extra_info();
+            auto it = extra.find("input_type");
+            if (it != extra.end())
+            {
+                if (it->second == "bgr")
+                    _input_img_fmt = image::Format::FMT_BGR888;
+                else if (it->second == "gray")
+                    _input_img_fmt = image::Format::FMT_GRAYSCALE;
+                else
+                    _input_img_fmt = image::Format::FMT_RGB888;
+            }
+            return err::ERR_NONE;
+        }
+
+        /**
+         * Get model input size, only for image input
+         * @return model input size
+         * @maixpy maix.nn.Yolo26Depth.input_size
+         */
+        image::Size input_size()
+        {
+            return image::Size(_input_w, _input_h);
+        }
+
+        /**
+         * Get model input width, only for image input
+         * @return model input size of width
+         * @maixpy maix.nn.Yolo26Depth.input_width
+         */
+        int input_width()
+        {
+            return _input_w;
+        }
+
+        /**
+         * Get model input height, only for image input
+         * @return model input size of height
+         * @maixpy maix.nn.Yolo26Depth.input_height
+         */
+        int input_height()
+        {
+            return _input_h;
+        }
+
+        /**
+         * Get input image format, only for image input
+         * @return input image format, image::Format type.
+         * @maixpy maix.nn.Yolo26Depth.input_format
+         */
+        image::Format input_format()
+        {
+            return _input_img_fmt;
+        }
+
+        /**
+         * Get model output size (depth map size), only for image input
+         * @return model output size
+         * @maixpy maix.nn.Yolo26Depth.output_size
+         */
+        image::Size output_size()
+        {
+            return image::Size(_output_w, _output_h);
+        }
+
+        /**
+         * Forward model and get raw image depth estimation data.
+         * @param img image, format should match model input_type， or will raise err.Exception
+         * @param fit image resize fit mode if input image not equal to model' input size,
+         *            will auto resize to model's input size then detect, and recover to image input size.
+         *            Default Fit.FIT_CONTAIN, see image.Fit.
+         * @throw If error occurred, will raise err::Exception, you can find reason in log, mostly caused by args error or hardware error.
+         * @return result, a tensor.Tensor object. If in dual_buff mode, value can be None(in Python) or nullptr(in C++) when not ready. In C++, you need to delete it after use.
+         * @maixpy maix.nn.Yolo26Depth.get_depth
+         */
+        tensor::Tensor *get_depth(image::Image &img, image::Fit fit = image::FIT_CONTAIN)
+        {
+            if (_model == nullptr) return nullptr;
+            tensor::Tensors *outputs = _model->forward_image(img, {}, {}, fit, false, false);
+            if (!outputs) return nullptr;
+            tensor::Tensor *t = outputs->begin()->second;
+            if (t->dtype() != tensor::DType::FLOAT32)
+            {
+                delete outputs;
+                return nullptr;
+            }
+            _output_h = t->shape()[2];
+            _output_w = t->shape()[3];
+            tensor::Tensor *result = new tensor::Tensor(t->shape(), t->dtype(), t->data(), true);
+            delete outputs;
+            return result;
+        }
+
+        /**
+         * Forward model and get image depth estimation data normlized to [0, 255] and as a image.Image object.
+         * @param img image, format should match model input_type， or will raise err.Exception
+         * @param fit image resize fit mode if input image not equal to model' input size,
+         *            will auto resize to model's input size then detect, and recover to image input size.
+         *            Default Fit.FIT_CONTAIN, see image.Fit.
+         * @param cmap Color map used convert grayscale distance estimation image to RGB image.
+         *             Diiferent cmap will influence finally image.
+         *             Default image.CMap.JET (near red/yellow, far blue).
+         * @throw If error occurred, will raise err::Exception, you can find reason in log, mostly caused by args error or hardware error.
+         * @return result, a image::Image object. If in dual_buff mode, value can be None(in Python) or nullptr(in C++) when not ready. In C++, you need to delete it after use.
+         * @maixpy maix.nn.Yolo26Depth.get_depth_image
+         */
+        image::Image *get_depth_image(image::Image &img, image::Fit fit = image::FIT_CONTAIN,
+                                      image::CMap cmap = image::CMap::JET)
+        {
+            if (_model == nullptr) return nullptr;
+            _cmap = cmap;
+
+            tensor::Tensors *outputs = _model->forward_image(img, {}, {}, fit, false, false);
+            if (!outputs) return nullptr;
+            tensor::Tensor *t = outputs->begin()->second;
+            if (t->dtype() != tensor::DType::FLOAT32)
+            {
+                delete outputs;
+                return nullptr;
+            }
+            int out_w = t->shape()[3];
+            int out_h = t->shape()[2];
+            _output_w = out_w;
+            _output_h = out_h;
+            const float *depth = (const float *)t->data();
+
+            // compute letterbox content region (remove padding), keep depth pixel-aligned with image
+            int top = 0, bottom = 0, left = 0, right = 0;
+            if (fit == image::Fit::FIT_CONTAIN && (img.width() != out_w || img.height() != out_h))
+            {
+                float gain = std::min((float)out_h / img.height(), (float)out_w / img.width());
+                int rw = (int)std::round(img.width() * gain);
+                int rh = (int)std::round(img.height() * gain);
+                int pw = out_w - rw;
+                int ph = out_h - rh;
+                left = (int)std::round(pw / 2.0f - 0.1f);
+                right = (int)std::round(pw / 2.0f + 0.1f);
+                top = (int)std::round(ph / 2.0f - 0.1f);
+                bottom = (int)std::round(ph / 2.0f + 0.1f);
+            }
+            int crop_w = out_w - left - right;
+            int crop_h = out_h - top - bottom;
+
+            // extract content region depth (row-major)
+            std::vector<float> crop(crop_w * crop_h);
+            for (int y = 0; y < crop_h; y++)
+                for (int x = 0; x < crop_w; x++)
+                    crop[y * crop_w + x] = depth[(y + top) * out_w + (x + left)];
+
+            // generate heatmap on crop size
+            image::Image *heatmap = _colorize(crop.data(), crop_w, crop_h);
+
+            // resize back to input image size
+            if (crop_w != img.width() || crop_h != img.height())
+            {
+                image::Image *result = heatmap->resize(img.width(), img.height(), image::FIT_FILL);
+                delete heatmap;
+                delete outputs;
+                return result;
+            }
+            delete outputs;
+            return heatmap;
+        }
+
+    private:
+        nn::NN *_model;
+        bool _dual_buff;
+        image::Format _input_img_fmt;
+        int _input_w, _input_h;
+        int _output_w, _output_h;
+        image::CMap _cmap;
+
+        /**
+         * Convert depth map to RGB888 heatmap: disparity(1/depth) + min-max normalize + cmap lookup.
+         * Return a new image::Image, caller should delete it.
+         */
+        image::Image *_colorize(const float *depth, int w, int h)
+        {
+            auto &colors = image::cmap_colors_rgb(_cmap);
+            image::Image *result = new image::Image(w, h, image::Format::FMT_RGB888);
+            uint8_t *img_data = (uint8_t *)result->data();
+            int n = w * h;
+
+            float min_v = FLT_MAX, max_v = -FLT_MAX;
+            for (int i = 0; i < n; i++)
+            {
+                float d = depth[i];
+                if (std::isfinite(d) && d > 0)
+                {
+                    float v = 1.0f / d;   // disparity
+                    if (v < min_v) min_v = v;
+                    if (v > max_v) max_v = v;
+                }
+            }
+            if (min_v == max_v)
+            {
+                memset(img_data, 127, n * 3);
+                return result;
+            }
+            float scale = 255.0f / (max_v - min_v);
+
+#pragma omp parallel for
+            for (int i = 0; i < n; i++)
+            {
+                float d = depth[i];
+                if (std::isfinite(d) && d > 0)
+                {
+                    uint8_t gray = (uint8_t)std::clamp((1.0f / d - min_v) * scale, 0.0f, 255.0f);
+                    const auto &rgb = colors[gray];
+                    img_data[3 * i + 0] = rgb[0];
+                    img_data[3 * i + 1] = rgb[1];
+                    img_data[3 * i + 2] = rgb[2];
+                }
+                else
+                {
+                    img_data[3 * i + 0] = 0;
+                    img_data[3 * i + 1] = 0;
+                    img_data[3 * i + 2] = 0;
+                }
+            }
+            return result;
+        }
+    };
+}
diff --git a/examples/nn_yolo26_depth/.gitignore b/examples/nn_yolo26_depth/.gitignore
new file mode 100644
index 000000000..777259b5c
--- /dev/null
+++ b/examples/nn_yolo26_depth/.gitignore
@@ -0,0 +1,5 @@
+build
+dist
+.config.mk
+.flash.conf.json
+data
diff --git a/examples/nn_yolo26_depth/README.md b/examples/nn_yolo26_depth/README.md
new file mode 100644
index 000000000..3c378ed2d
--- /dev/null
+++ b/examples/nn_yolo26_depth/README.md
@@ -0,0 +1,35 @@
+# YOLO26-depth example
+
+YOLO26-depth (monocular depth estimation) example for MaixCDK, based on the header-only class `maix_nn_yolo26_depth.hpp`.
+
+## Build
+
+```bash
+cd examples/nn_yolo26_depth
+export CMAKE_POLICY_VERSION_MINIMUM=3.5   # required on CMake >= 4.x, see FAQ
+maixcdk build -p maixcam2
+```
+
+## Usage
+
+Single image inference, save heatmap:
+
+```bash
+./nn_yolo26_depth /tmp/yolo26n-depth.mud /tmp/bus.jpg /tmp/bus_heatmap.jpg
+```
+
+Real-time camera inference, display heatmap on screen (exit with SIGINT):
+
+```bash
+./nn_yolo26_depth /tmp/yolo26n-depth.mud
+```
+
+## Model files
+
+Put the model on the device first:
+
+```bash
+scp yolo26n-depth.mud yolo26n-depth_w8a8_mix.axmodel root@<device-ip>:/tmp/
+```
+
+Build method please visit [MaixCDK](https://github.com/sipeed/MaixCDK).
diff --git a/examples/nn_yolo26_depth/main/CMakeLists.txt b/examples/nn_yolo26_depth/main/CMakeLists.txt
new file mode 100644
index 000000000..6619a51c7
--- /dev/null
+++ b/examples/nn_yolo26_depth/main/CMakeLists.txt
@@ -0,0 +1,16 @@
+############### Add include ###################
+list(APPEND ADD_INCLUDE "include"
+    )
+list(APPEND ADD_PRIVATE_INCLUDE "")
+###############################################
+
+############ Add source files #################
+append_srcs_dir(ADD_SRCS "src")       # append source file in src dir to var ADD_SRCS
+###############################################
+
+###### Add required/dependent components ######
+list(APPEND ADD_REQUIREMENTS basic nn vision)
+###############################################
+
+# register component, DYNAMIC or SHARED flags will make component compiled to dynamic(shared) lib
+register_component()
diff --git a/examples/nn_yolo26_depth/main/Kconfig b/examples/nn_yolo26_depth/main/Kconfig
new file mode 100644
index 000000000..0ee4d182c
--- /dev/null
+++ b/examples/nn_yolo26_depth/main/Kconfig
@@ -0,0 +1,9 @@
+menu "Main Component Config"
+    config MAIN_COMPONENT_EACH_BUILD
+        bool "Each component build a separate shared library file"
+        default n
+
+    config MAIN_COMPONENT_BUILD_TYPE
+        string "Main component build type"
+        default "Release"
+endmenu
diff --git a/examples/nn_yolo26_depth/main/include/main.h b/examples/nn_yolo26_depth/main/include/main.h
new file mode 100644
index 000000000..ff6398284
--- /dev/null
+++ b/examples/nn_yolo26_depth/main/include/main.h
@@ -0,0 +1,3 @@
+#pragma once
+
+#define APP_VERSION "0.1.0"
diff --git a/examples/nn_yolo26_depth/main/src/main.cpp b/examples/nn_yolo26_depth/main/src/main.cpp
new file mode 100644
index 000000000..40f174c5b
--- /dev/null
+++ b/examples/nn_yolo26_depth/main/src/main.cpp
@@ -0,0 +1,134 @@
+/**
+ * YOLO26-depth example
+ *
+ * Usage:
+ *   - 单张图片推理并保存热力图:
+ *       ./nn_yolo26_depth mud_model_path image_path [output_path]
+ *       例: ./nn_yolo26_depth /tmp/yolo26n-depth.mud /tmp/bus.jpg /tmp/bus_heatmap.jpg
+ *   - 实时摄像头推理并显示热力图:
+ *       ./nn_yolo26_depth mud_model_path
+ *
+ * 依赖(设备上需存在):
+ *   /tmp/yolo26n-depth.mud
+ *   /tmp/yolo26n-depth_w8a8_mix.axmodel
+ *
+ * 退出: 摄像头模式下发送 SIGINT/SIGTERM(kill -INT <pid>) 或设备退出信号。
+ */
+
+#include "maix_basic.hpp"
+#include "maix_camera.hpp"
+#include "maix_display.hpp"
+#include "maix_image.hpp"
+#include "maix_nn_yolo26_depth.hpp"
+#include "maix_time.hpp"
+#include "main.h"
+
+using namespace maix;
+
+int _main(int argc, char *argv[])
+{
+    int ret = 0;
+    log::info("Program start");
+    std::string help = "Usage: " + std::string(argv[0]) + " mud_model_path [image_path [output_path]]";
+
+    if (argc < 2)
+    {
+        log::error(help.c_str());
+        return -1;
+    }
+
+    const char *model_path = argv[1];
+    bool dual_buff = false;
+    image::CMap cmap = image::CMap::JET;
+
+    log::info("model path: %s", model_path);
+    nn::Yolo26Depth model(model_path, dual_buff);
+    log::info("load model %s success", model_path);
+    log::info("model input size: %dx%d, format: %d",
+              model.input_width(), model.input_height(), (int)model.input_format());
+    log::info("dual buff mode: %d", dual_buff);
+
+    if (argc >= 3)
+    {
+        // ---------- 模式一: 单张图片推理, 保存热力图 ----------
+        const char *img_path = argv[2];
+        std::string output_path = (argc >= 4) ? argv[3] : "depth_heatmap.jpg";
+        log::info("load image now");
+        maix::image::Image *img = maix::image::load(img_path, model.input_format());
+        err::check_null_raise(img, "load image " + std::string(img_path) + " failed");
+        log::info("load image %s success: %s", img_path, img->to_str().c_str());
+        if (img->width() != model.input_width() || img->height() != model.input_height())
+        {
+            log::warn("image size not match model input size, will auto resize from %dx%d to %dx%d",
+                      img->width(), img->height(), model.input_width(), model.input_height());
+        }
+
+        // 推理得到深度热力图(自动 resize 回原图尺寸)
+        maix::image::Image *heatmap = model.get_depth_image(*img, image::FIT_CONTAIN, cmap);
+        err::check_null_raise(heatmap, "get_depth_image failed");
+
+        // 保存热力图
+        err::Err e = heatmap->save(output_path);
+        if (e != err::ERR_NONE)
+        {
+            log::error("save heatmap to %s failed", output_path.c_str());
+            ret = -1;
+        }
+        else
+        {
+            log::info("save heatmap to %s success", output_path.c_str());
+        }
+
+        delete heatmap;
+        delete img;
+    }
+    else
+    {
+        // ---------- 模式二: 实时摄像头推理并显示热力图 ----------
+        camera::Camera cam = camera::Camera(model.input_width(), model.input_height(), model.input_format());
+        display::Display disp = display::Display();
+        log::info("camera %dx%d", cam.width(), cam.height());
+
+        uint64_t frame_id = 0;
+        uint64_t t_last = time::ticks_ms();
+        int fps = 0;
+
+        while (!app::need_exit())
+        {
+            uint64_t t0 = time::ticks_ms();
+            image::Image *img = cam.read();
+            if (!img)
+            {
+                time::sleep_ms(10);
+                continue;
+            }
+
+            image::Image *heatmap = model.get_depth_image(*img, image::FIT_CONTAIN, cmap);
+            if (heatmap)
+            {
+                disp.show(*heatmap);
+                delete heatmap;
+            }
+            delete img;
+
+            frame_id++;
+            if (frame_id % 20 == 0)
+            {
+                uint64_t now = time::ticks_ms();
+                fps = (int)(20000.0f / (now - t_last));
+                t_last = now;
+                log::info("frame=%llu e2e=%llu ms fps=%d", (unsigned long long)frame_id,
+                          (unsigned long long)(time::ticks_ms() - t0), fps);
+            }
+        }
+    }
+
+    log::info("Program exit");
+    return ret;
+}
+
+int main(int argc, char *argv[])
+{
+    sys::register_default_signal_handle();
+    CATCH_EXCEPTION_RUN_RETURN(_main, -1, argc, argv);
+}