# Inquiry About YOLO26n Deployment on AMB82-Mini Using Custom Model Conversion

**URL:** <https://forum.amebaiot.com/t/inquiry-about-yolo26n-deployment-on-amb82-mini-using-custom-model-conversion/5072>\
**Category:** SDK\
**Tags:** evb-amb82\
**Created:** [September 18, 2026, 1:57am UTC](https://forum.amebaiot.com/t/inquiry-about-yolo26n-deployment-on-amb82-mini-using-custom-model-conversion/5072 "2026-09-18T01:57:41Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![Eranga\_Wijebandara](https://sea2.discourse-cdn.com/flex016/user_avatar/forum.amebaiot.com/eranga_wijebandara/32/3529_2.png) [@Eranga\_Wijebandara](https://forum.amebaiot.com/u/Eranga_Wijebandara)\
**Post date:** [September 18, 2026, 1:57am UTC](https://forum.amebaiot.com/t/inquiry-about-yolo26n-deployment-on-amb82-mini-using-custom-model-conversion/5072/1 "2026-09-18T01:57:41Z")

</div>

### Dear Realtek AI Support Team,

I am currently working on an AI object detection project using the **Realtek AMB82-Mini development board**. I would like to train and deploy the YOLO26n model for real-time object detection.

I reviewed your model conversion instructions, which mention TensorFlow versions up to 2.14.1 and saving models in `.h5` format using `include_optimizer=False`.

However, I have some questions regarding YOLO26n compatibility.

**My questions are:**

1. Does the Realtek AI model conversion tool support **YOLO26n** for deployment on the AMB82-Mini?

2. Can I train YOLO26n using Google Colab and convert the trained model into a Realtek-compatible `.nb` file?

3. Is it possible to convert YOLO26n into TensorFlow/Keras `.h5` format and submit it to your conversion service?

4. Does the AMB82-Mini support running YOLO26n using the **0.4 TOPS NPU**?

5. Is there an offline model conversion tool available that supports custom YOLO architectures?

6. If YOLO26n is not currently supported, are there any recommended methods for deploying custom YOLO models without using the Realtek conversion tool?

7. Is there any documentation or example available for deploying custom object detection models using TensorFlow Lite or another supported runtime?

My intended workflow is:

```auto
Google Colab
    ↓
Train YOLO26n
    ↓
Export trained model
    ↓
Realtek model conversion
    ↓
Deploy on AMB82-Mini
    ↓
Real-time camera object detection

```

I would appreciate your guidance on the supported model formats, conversion requirements, and recommended deployment process.

Thank you for your support.

---

<div class="post-metadata">

**Author:** ![system](https://us1.discourse-cdn.com/flex016/uploads/ameba/original/2X/7/7009caf971627f868f3de3d34312aff2d14c61e3.png) [@system](https://forum.amebaiot.com/u/system)\
**Post date:** [September 18, 2026, 1:57am UTC](https://forum.amebaiot.com/t/inquiry-about-yolo26n-deployment-on-amb82-mini-using-custom-model-conversion/5072/2 "2026-09-18T01:57:50Z")

</div>

👋 Thanks for posting!

For documentation, SDK resources, FAQs, and community guidelines, please visit: [here](https://forum.amebaiot.com/t/forum-community-guidelines-and-resources/4969)

For commercial users, if you would like to have dedicated technical support or to connect with us, please fill in the [Private User Form](https://forms.cloud.microsoft/r/FK9WxDhvDg)

Happy building with Ameba!

* * *

👋 感謝您的分享！

如需查閱技術文件、SDK 資源、常見問題及社群指南，請參考: [這裡](https://forum.amebaiot.com/t/forum-community-guidelines-and-resources/4969)

若您是商業用戶，希望獲得專屬技術支援或與我們聯繫，請填寫 [Private User Form (商業用戶表單)](https://forms.cloud.microsoft/r/FK9WxDhvDg)

祝您使用 Ameba 開發順利！

---

<div class="post-metadata">

**Author:** ![KevinKL](https://sea2.discourse-cdn.com/flex016/user_avatar/forum.amebaiot.com/kevinkl/32/2632_2.png) [@KevinKL](https://forum.amebaiot.com/u/KevinKL)\
**Post date:** [September 30, 2026, 9:45am UTC](https://forum.amebaiot.com/t/inquiry-about-yolo26n-deployment-on-amb82-mini-using-custom-model-conversion/5072/3 "2026-09-30T09:45:46Z")

</div>

Hi @Eranga_Wijebandara ,

We want to raise a critical technical consideration: the current AMB82-mini SDK’s native post-processing pipeline currently supports YOLOv3-tiny, YOLOv4-tiny, and YOLOv7-tiny for object detection. YOLOv26n is not natively supported in the current SDK.

Additionally, please be aware:

- The Pro2 NPU does not support per-channel quantized models — only per-tensor quantization is supported.

- For PyTorch-based models (including YOLOv26), we strongly recommend exporting to ONNX first before importing into the Acuity Toolkit.

Our recommendation for your project is to consider retraining with YOLOv7-tiny. YOLOv7-tiny has full SDK support including post-processing, reparameterization scripts, and deployment examples. If you wish to proceed with YOLOv26n, be advised that you would need to write custom post-processing code on the device side, as the output tensor format of YOLOv26 differs from the YOLOv3/4/7 family that the SDK natively handles.

Thank you.