I downloaded the NeuroPilot SDK from
In the extracted folders, I see several directories (such as host, mt6878, mt6897, mt6899, mt6989, and mt6991), each containing both neuronrt and ncc-tflite files.
How do I choose the correct versions for deploying a model on Genio-720?
1 Like
Thank you for this question.
The MediaTek Genio-720 /Genio-520 correspond to the mt8189 code name in the NeuroPilot ecosystem.
mt8189 belongs to the Genio IoT series and uses MDLA5.3 (MediaTek Deep Learning Accelerator version 5.3) with NeuroPilot 8 (NP8) support.
Important:
As of now, the official NeuroPilot SDK All-In-One Bundle does not include binaries for mt8189, which are required for Genio-720/Genio-520 deployments.
Recommended Download for Genio-720
Please download the dedicated mt8189 package from:
Version Table: Mapping SoC Code Names to IoT Series and MDLA Versions
| Directory Name |
IoT Code Name |
IoT Product Series |
NeuroPilot Version |
MDLA Version |
| mt6878 |
8873 |
IoT-Modem |
NP7 |
MDLA3.5 |
| mt6897 |
8883 |
IoT-Modem |
NP7 |
MDLA5.0 |
| mt6989 |
8893 |
IoT-Modem |
NP7 |
MDLA5.1 |
| mt6991 |
8894 |
IoT-Modem |
NP8 |
MDLA5.5 |
| mt8189 |
8391 |
IoT-Genio |
NP8 |
MDLA5.3 |
Tool Selection: neuronrt and ncc-tflite Usage Guide
-
neuronrt
- In the
host folder: Early developer validation tool, intended for PC environments with compatible MediaTek NPU. Not used for on-device deployment.
- In device-specific folders (e.g.,
mt8189): Validation and inference tool for on-device usage—select the version matching your device’s SoC code name.
-
ncc-tflite
- Host tool for converting TFLite models into DLA binaries targeting a specific hardware generation (e.g., MDLA3, MDLA5).
- Use the version that corresponds to your deployment target (e.g., for Genio-720 use the one from the
mt8189 package).
Deployment Steps for Genio-720 / Genio-520
- Download the mt8189-specific package from the NeuroPilot download page (not the All-In-One Bundle).
- Use the
ncc-tflite tool from the mt8189 package to convert TFLite models to DLA binaries targeting MDLA5.3.
- Use the
neuronrt tool from the mt8189 package on your Genio-720 device for model validation or on-device inference.
MediaTek Recommendation: Prefer Host Tool for Model Conversion
MediaTek strongly recommends using the host version of ncc-tflite (located in the host folder) for model conversion tasks. The host-side ncc-tflite tool provides:
-
More comprehensive operator (op) support than device-specific versions
-
Improved memory management during model conversion
-
Better overall reliability, reducing the chance of memory allocation failures that can sometimes arise with device-side tools
Always perform model conversion on your development host PC with the host ncc-tflite, then transfer the generated DLA binaries for deployment on the target device.
Summary
- Do not use the
neuronrt or ncc-tflite tools from other SoC folders (e.g., mt6878, mt6991) for Genio-720.
- Always match the tool version to your target device’s code name (mt8189 for Genio-720/Genio-520).
- For complete documentation and updates, refer to the official MediaTek NeuroPilot download page.
1 Like