MNASNet05: Optimized for Qualcomm Devices

MNASNet05 is a machine learning model that can classify images from the Imagenet dataset. It can also be used as a backbone in building more complex models for specific use cases.

This is based on the implementation of MNASNet05 found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.

Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.

Getting Started

There are two ways to deploy this model on your device:

Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

Runtime	Precision	Chipset	SDK Versions	Download
ONNX	float	Universal	QAIRT 2.42, ONNX Runtime 1.24.3	Download
ONNX	w8a16	Universal	QAIRT 2.42, ONNX Runtime 1.24.3	Download
QNN_DLC	float	Universal	QAIRT 2.43	Download
QNN_DLC	w8a16	Universal	QAIRT 2.43	Download
TFLITE	float	Universal	QAIRT 2.43, TFLite 2.19.1	Download

For more device-specific assets and performance metrics, visit MNASNet05 on Qualcomm® AI Hub.

Option 2: Export with Custom Configurations

Use the Qualcomm® AI Hub Models Python library to compile and export the model with your own:

• Custom weights (e.g., fine-tuned checkpoints)
• Custom input shapes
• Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for MNASNet05 on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.image_classification

Model Stats:

• Model checkpoint: Imagenet
• Input resolution: 224x224
• Number of parameters: 2.21M
• Model size (float): 8.45 MB
• Model size (w8a16): 2.79 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
MNASNet05	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	0.218 ms	0 - 31 MB	NPU
MNASNet05	ONNX	float	Snapdragon® X2 Elite	0.251 ms	5 - 5 MB	NPU
MNASNet05	ONNX	float	Snapdragon® X Elite	0.611 ms	5 - 5 MB	NPU
MNASNet05	ONNX	float	Snapdragon® 8 Gen 3 Mobile	0.339 ms	0 - 47 MB	NPU
MNASNet05	ONNX	float	Qualcomm® QCS8550 (Proxy)	0.486 ms	1 - 3 MB	NPU
MNASNet05	ONNX	float	Qualcomm® QCS9075	0.763 ms	1 - 3 MB	NPU
MNASNet05	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	0.265 ms	0 - 27 MB	NPU
MNASNet05	ONNX	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	0.218 ms	0 - 33 MB	NPU
MNASNet05	ONNX	w8a16	Snapdragon® X2 Elite	0.22 ms	0 - 0 MB	NPU
MNASNet05	ONNX	w8a16	Snapdragon® X Elite	0.64 ms	2 - 2 MB	NPU
MNASNet05	ONNX	w8a16	Snapdragon® 8 Gen 3 Mobile	0.34 ms	0 - 40 MB	NPU
MNASNet05	ONNX	w8a16	Qualcomm® QCS6490	29.339 ms	10 - 13 MB	CPU
MNASNet05	ONNX	w8a16	Qualcomm® QCS8550 (Proxy)	0.517 ms	0 - 8 MB	NPU
MNASNet05	ONNX	w8a16	Qualcomm® QCS9075	0.69 ms	0 - 3 MB	NPU
MNASNet05	ONNX	w8a16	Qualcomm® QCM6690	10.32 ms	9 - 16 MB	CPU
MNASNet05	ONNX	w8a16	Snapdragon® 8 Elite For Galaxy Mobile	0.261 ms	0 - 33 MB	NPU
MNASNet05	ONNX	w8a16	Snapdragon® 7 Gen 4 Mobile	7.63 ms	10 - 17 MB	CPU
MNASNet05	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	0.301 ms	1 - 33 MB	NPU
MNASNet05	QNN_DLC	float	Snapdragon® X2 Elite	0.438 ms	1 - 1 MB	NPU
MNASNet05	QNN_DLC	float	Snapdragon® X Elite	0.965 ms	1 - 1 MB	NPU
MNASNet05	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	0.524 ms	0 - 46 MB	NPU
MNASNet05	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	2.322 ms	1 - 29 MB	NPU
MNASNet05	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	0.821 ms	1 - 9 MB	NPU
MNASNet05	QNN_DLC	float	Qualcomm® SA8775P	1.119 ms	1 - 31 MB	NPU
MNASNet05	QNN_DLC	float	Qualcomm® QCS9075	0.976 ms	1 - 3 MB	NPU
MNASNet05	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	1.575 ms	0 - 48 MB	NPU
MNASNet05	QNN_DLC	float	Qualcomm® SA7255P	2.322 ms	1 - 29 MB	NPU
MNASNet05	QNN_DLC	float	Qualcomm® SA8295P	1.429 ms	0 - 28 MB	NPU
MNASNet05	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	0.401 ms	1 - 34 MB	NPU
MNASNet05	QNN_DLC	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	0.297 ms	0 - 29 MB	NPU
MNASNet05	QNN_DLC	w8a16	Snapdragon® X2 Elite	0.417 ms	0 - 0 MB	NPU
MNASNet05	QNN_DLC	w8a16	Snapdragon® X Elite	0.935 ms	0 - 0 MB	NPU
MNASNet05	QNN_DLC	w8a16	Snapdragon® 8 Gen 3 Mobile	0.533 ms	0 - 37 MB	NPU
MNASNet05	QNN_DLC	w8a16	Qualcomm® QCS6490	2.243 ms	2 - 4 MB	NPU
MNASNet05	QNN_DLC	w8a16	Qualcomm® QCS8275 (Proxy)	1.653 ms	0 - 25 MB	NPU
MNASNet05	QNN_DLC	w8a16	Qualcomm® QCS8550 (Proxy)	0.787 ms	0 - 2 MB	NPU
MNASNet05	QNN_DLC	w8a16	Qualcomm® SA8775P	0.994 ms	0 - 27 MB	NPU
MNASNet05	QNN_DLC	w8a16	Qualcomm® QCS9075	0.938 ms	0 - 2 MB	NPU
MNASNet05	QNN_DLC	w8a16	Qualcomm® QCM6690	3.042 ms	0 - 139 MB	NPU
MNASNet05	QNN_DLC	w8a16	Qualcomm® QCS8450 (Proxy)	0.954 ms	0 - 38 MB	NPU
MNASNet05	QNN_DLC	w8a16	Qualcomm® SA7255P	1.653 ms	0 - 25 MB	NPU
MNASNet05	QNN_DLC	w8a16	Qualcomm® SA8295P	1.314 ms	0 - 23 MB	NPU
MNASNet05	QNN_DLC	w8a16	Snapdragon® 8 Elite For Galaxy Mobile	0.358 ms	0 - 29 MB	NPU
MNASNet05	QNN_DLC	w8a16	Snapdragon® 7 Gen 4 Mobile	0.796 ms	0 - 24 MB	NPU
MNASNet05	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	0.307 ms	0 - 34 MB	NPU
MNASNet05	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	0.528 ms	0 - 48 MB	NPU
MNASNet05	TFLITE	float	Qualcomm® QCS8275 (Proxy)	2.352 ms	0 - 30 MB	NPU
MNASNet05	TFLITE	float	Qualcomm® QCS8550 (Proxy)	0.802 ms	0 - 2 MB	NPU
MNASNet05	TFLITE	float	Qualcomm® SA8775P	1.15 ms	0 - 32 MB	NPU
MNASNet05	TFLITE	float	Qualcomm® QCS9075	0.986 ms	0 - 8 MB	NPU
MNASNet05	TFLITE	float	Qualcomm® QCS8450 (Proxy)	1.581 ms	0 - 49 MB	NPU
MNASNet05	TFLITE	float	Qualcomm® SA7255P	2.352 ms	0 - 30 MB	NPU
MNASNet05	TFLITE	float	Qualcomm® SA8295P	1.46 ms	0 - 29 MB	NPU
MNASNet05	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	0.397 ms	0 - 30 MB	NPU

License

• The license for the original implementation of MNASNet05 can be found here.

References

• MnasNet: Platform-Aware Neural Architecture Search for Mobile
• Source Model Implementation

Community

• Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
• For questions or feedback please reach out to us.