By clicking “Accept All Cookies”, you agree to the storing of cookies on your device to enhance site navigation, analyze site usage and assist in our marketing efforts. More info
Cookie Settings
By clicking “Accept All Cookies”, you agree to the storing of cookies on your device to enhance site navigation, analyze site usage and assist in our marketing efforts. More info
Nowadays, AI innovations have expanded beyond remote server environments. AI solutions are being integrated into edge devices such as sensors and smartphones. Thanks to this technological shift, data can now be handled directly where it is generated, enabling faster responses, improved privacy, and reduced reliance on constant cloud connectivity.
As a result, edge AI is gaining traction across many industries. The edge AI software market is expected to reach $8.88 billion by 2031 as more systems move toward faster and more local processing.
In particular, computer vision, a branch of AI that focuses on understanding images and video, is being rapidly adopted at the edge. From counting food items as they’re packaged to helping vehicles detect pedestrians, computer vision supports countless practical applications across different sectors.
This is made possible through computer vision models. For instance, Ultralytics YOLO11 is a model that supports various Vision AI tasks like object detection, instance segmentation, object tracking, and pose estimation. It is designed to be fast and efficient and performs well on devices with limited hardware resources.
Fig 1. Detecting and tracking food being packaged using YOLO11 (Source).
In addition to being suitable for edge deployment, through various integrations supported by Ultralytics, YOLO11 can be exported to various formats suitable for different hardware environments.
One of the most efficient options is MNN (Mobile Neural Network), a lightweight inference engine designed for low-resource devices. Exporting YOLO11 to MNN enables it to run directly on mobile phones, embedded systems, and other edge platforms where fast, on-device processing is essential.
In this article, we’ll explore how the MNN integration works, highlight common use cases, and walk through how to get started with running inferences using an exported YOLO11 model. Let’s get started!
An overview of MNN: A deep learning framework
Running computer vision models on smaller devices like mobile phones, industrial sensors, and portable systems isn’t always straightforward. These devices often have limited memory, slower processors, and strict power limits.
The Mobile Neural Network, or MNN, is a lightweight and high-performance inference engine developed by Alibaba to make AI models run efficiently on low-resource hardware while maintaining real-time performance. MNN supports a wide range of platforms, including Android, iOS, and Linux, and works across a range of hardware types like central processing units (CPUs) and graphics processing units (GPUs).
The MNN integration supported by Ultralytics makes it possible to export YOLO11 models easily to the MNN format. Simply put, this means the models can be converted from the YOLO format to MNN.
Once converted, they can be deployed on devices that support the MNN framework for efficient, on-device inference. A key benefit of using the MNN format is that it simplifies deploying YOLO11 in scenarios where size, speed, and resource efficiency are critical.
Key features of the MNN inference backend
Before we dive into how to use the MNN integration, let's take a look at what makes the MNN framework a great choice for running AI models on real-world devices. It’s built to handle the unique constraints of edge environments while still delivering fast and reliable performance.
Interestingly, MNN is used internally at Alibaba in over 30 applications, including Taobao, Tmall, Youku, DingTalk, and Xianyu, across a wide range of scenarios like live video, short-form content, image search, and on-device security checks. It supports large-scale deployment and runs millions of inferences per day in production environments.
Here are some of the key features of the MNN framework:
Backend auto-selection: MNN can automatically choose the most suitable execution backend, such as CPU or GPU, based on the hardware it’s running on.
Multi-threaded execution: It supports multi-threading, allowing it to take full advantage of multicore processors for faster inference.
Supports model quantization: It enables you to reduce model size significantly using FP16 or INT8 quantization, helping improve inference speed while using less memory.
Lightweight and fast: MNN has a very small footprint, with the core library around 400 KB on Android and about 5 MB on iOS, which makes it ideal for mobile and embedded devices.
Understanding how the MNN integration works
Next, let’s walk through how to export YOLO11 models to the MNN format.
The first step is to install the Ultralytics Python package, which provides everything needed to export YOLO11 models to the MNN format. You can do this by running "pip install ultralytics" on your terminal or using the command prompt. If you're using a Jupyter Notebook or Google Colab, add an exclamation mark before the command.
If you encounter any issues during installation, refer to the Common Issues guide for troubleshooting tips.
Once your environment is set up, you can load a pre-trained YOLO11 model such as "yolo11n.pt" and export it to the MNN format as shown in the code snippet below. If you’ve trained your own custom YOLO11 model, you can export it simply by replacing the filename with your model's path.
from ultralytics import YOLO
model = YOLO("yolo11n.pt")
model.export(format="mnn")
After converting your model to MNN, you can use it across different mobile and embedded platforms depending on your application needs.
For example, suppose you want to test the exported model on a video of traffic. In that case, you can load the YOLO11 model in MNN format to detect objects such as vehicles, pedestrians, and traffic signs directly on the device, as shown in the example below.
When the inference is complete, the output video with detected objects is saved automatically in the 'runs/detect/predict' folder. Also, if you want to run inference using the MNN Python package directly, you can check out the official Ultralytics documentation for more details and examples.
Fig 3. Analyzing traffic using a YOLO11 model exported to MNN format. Image by author.
Use cases of edge AI model deployment enabled by YOLO11 and MNN
Deploying YOLO11 with MNN enables fast, efficient computer vision tasks such as object detection in environments where cloud-based processing isn’t practical or possible. Let’s see how this integration can be especially useful in real-world scenarios.
Mobile edge AI for plant disease identification
Plant disease identification apps that use image classification are gaining popularity among gardeners, researchers, and nature enthusiasts. With just a photo, users can quickly identify early signs of disease, such as leaf spots or discoloration. Since these apps are often used in outdoor areas where internet access may be limited or unavailable, relying on cloud processing can be unreliable.
After training, a YOLO11 model can be exported to the MNN format and run directly on mobile devices. The model can then classify plant species and detect visible disease symptoms locally, without sending any data to a server.
Fig 4. An example of using YOLO11 to detect signs of rust (a plant disease) on a leaf (Source).
Efficient on‑device inferences in manufacturing
Accurate package tracking is essential on busy production lines in manufacturing facilities. YOLO11 can be used to track and count each item as it moves through key checkpoints, updating counts in real time and flagging any discrepancies. This helps reduce missed or unaccounted shipments and supports smoother, more reliable operations.
Fig 5. Tracking and counting packages using YOLO11 (Source).
The MNN integration can be especially impactful in this context. Once the YOLO11 model is exported to MNN format, it can run directly on compact, low-power devices installed along the conveyor.
Because all processing happens locally, the system can deliver instant feedback and requires no internet connection. This ensures fast, reliable performance on the factory floor, keeping production moving efficiently while maintaining high accuracy and control.
Advantages of exporting YOLO11 to the MNN model format
Here are some key benefits of the MNN integration provided by Ultralytics:
Faster response times: Since inference runs on the device, predictions happen in real time with minimal latency.
Improved data privacy: Data stays on the device, reducing the need to send sensitive images or video to the cloud.
Open-source and actively maintained: Backed by Alibaba and supported by an active community, MNN is reliable and regularly updated with performance improvements.
Factors to consider when using the MNN framework
Before choosing MNN as your deployment framework, it's also important to evaluate how well it fits your project's requirements, deployment targets, and technical limitations. Here are some key factors to consider:
Ongoing compatibility: Framework updates or changes to your target platforms might require retesting or adjustments to keep everything running smoothly.
Fewer debugging tools: Compared to larger frameworks, MNN has more limited tools for debugging and inspecting model behavior, which can make troubleshooting harder.
Performance depends on hardware: The speed and efficiency of your model will vary depending on the device. Test your target hardware to be sure it meets your performance goals.
Key takeaways
Ultralytics’ support for MNN integration makes it easy to export YOLO11 models for use on mobile and embedded devices. It’s a practical option for applications that require fast, reliable detection without depending on cloud access or constant connectivity.
This setup helps streamline deployment while maintaining performance and keeping resource demands low. Whether you're building smart home systems, field tools, or compact industrial devices, exporting YOLO11 to MNN provides a flexible and efficient way to run computer vision tasks directly on edge devices.
EN
ZH
KO
JA
RU
DE
FR
ES
PT
IT
TR
VI
AR
.webp)
