OpenAI-Clip: Optimized for Qualcomm Devices

Contrastive Language-Image Pre-Training (CLIP) uses a ViT like transformer to get visual features and a causal language model to get the text features. Both the text and visual features can then be used for a variety of zero-shot learning tasks.

This is based on the implementation of OpenAI-Clip 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 2.45, ONNX Runtime 1.27.1	Download
ONNX	w8a16	Universal	QAIRT 2.45, ONNX Runtime 1.27.1	Download
QNN_DLC	float	Universal	QAIRT 2.45	Download
QNN_DLC	w8a16	Universal	QAIRT 2.45	Download
TFLITE	float	Universal	QAIRT 2.45	Download

For more device-specific assets and performance metrics, visit OpenAI-Clip 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 OpenAI-Clip on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.image_classification

Model Stats:

• Image input resolution: 224x224
• Model checkpoint: ViT-B/16
• Model size (float): 571 MB
• Number of parameters: 150M
• Text context length: 77

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
OpenAI-Clip	ONNX	float	Snapdragon® X2 Elite	15.041 ms	2 - 2 MB	NPU
OpenAI-Clip	ONNX	float	Snapdragon® X Elite	33.087 ms	294 - 294 MB	NPU
OpenAI-Clip	ONNX	float	Snapdragon® 8 Gen 3 Mobile	21.543 ms	0 - 1185 MB	NPU
OpenAI-Clip	ONNX	float	Snapdragon® 8 Gen 1 Mobile	47.884 ms	1 - 894 MB	NPU
OpenAI-Clip	ONNX	float	Qualcomm® Dragonwing™ IQ-8275	35.963 ms	0 - 4 MB	NPU
OpenAI-Clip	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	31.992 ms	1 - 4 MB	NPU
OpenAI-Clip	ONNX	float	Qualcomm® QCS8450	47.884 ms	1 - 894 MB	NPU
OpenAI-Clip	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	47.242 ms	0 - 4 MB	NPU
OpenAI-Clip	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	33.087 ms	294 - 294 MB	NPU
OpenAI-Clip	ONNX	float	Qualcomm® Dragonwing™ Q-8750	16.846 ms	1 - 958 MB	NPU
OpenAI-Clip	ONNX	float	Snapdragon® 8 Elite Mobile	16.846 ms	1 - 958 MB	NPU
OpenAI-Clip	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	13.451 ms	1 - 807 MB	NPU
OpenAI-Clip	ONNX	w8a16	Snapdragon® X2 Elite	13.395 ms	1 - 1 MB	NPU
OpenAI-Clip	ONNX	w8a16	Snapdragon® X Elite	27.143 ms	174 - 174 MB	NPU
OpenAI-Clip	ONNX	w8a16	Snapdragon® 8 Gen 3 Mobile	18.713 ms	0 - 750 MB	NPU
OpenAI-Clip	ONNX	w8a16	Snapdragon® 8 Gen 1 Mobile	39.887 ms	0 - 757 MB	NPU
OpenAI-Clip	ONNX	w8a16	Qualcomm® Dragonwing™ QCS6490	145.153 ms	1 - 4 MB	NPU
OpenAI-Clip	ONNX	w8a16	Qualcomm® Dragonwing™ IQ-8275	24.715 ms	0 - 4 MB	NPU
OpenAI-Clip	ONNX	w8a16	Qualcomm® Dragonwing™ QCS8550 (Proxy)	26.341 ms	0 - 4 MB	NPU
OpenAI-Clip	ONNX	w8a16	Qualcomm® QCS8450	39.887 ms	0 - 757 MB	NPU
OpenAI-Clip	ONNX	w8a16	Qualcomm® Dragonwing™ IQ-9075	28.04 ms	0 - 3 MB	NPU
OpenAI-Clip	ONNX	w8a16	Qualcomm® Dragonwing™ IQ-X7181	27.143 ms	174 - 174 MB	NPU
OpenAI-Clip	ONNX	w8a16	Qualcomm® Dragonwing™ Q-6690	303.476 ms	1 - 667 MB	NPU
OpenAI-Clip	ONNX	w8a16	Qualcomm® Dragonwing™ Q-7790	37.384 ms	1 - 670 MB	NPU
OpenAI-Clip	ONNX	w8a16	Qualcomm® Dragonwing™ Q-8750	14.885 ms	0 - 641 MB	NPU
OpenAI-Clip	ONNX	w8a16	Snapdragon® 8 Elite Mobile	14.885 ms	0 - 641 MB	NPU
OpenAI-Clip	ONNX	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	12.908 ms	0 - 684 MB	NPU
OpenAI-Clip	ONNX	w8a16	Snapdragon® 7 Gen 4 Mobile	37.384 ms	1 - 670 MB	NPU
OpenAI-Clip	QNN_DLC	float	Snapdragon® X2 Elite	14.657 ms	1 - 1 MB	NPU
OpenAI-Clip	QNN_DLC	float	Snapdragon® X Elite	33.277 ms	1 - 1 MB	NPU
OpenAI-Clip	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	21.772 ms	0 - 924 MB	NPU
OpenAI-Clip	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	46.455 ms	0 - 854 MB	NPU
OpenAI-Clip	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-8275	36.428 ms	1 - 3 MB	NPU
OpenAI-Clip	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-8275	98.999 ms	1 - 704 MB	NPU
OpenAI-Clip	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	31.465 ms	1 - 2 MB	NPU
OpenAI-Clip	QNN_DLC	float	Qualcomm® SA8775P	36.693 ms	1 - 701 MB	NPU
OpenAI-Clip	QNN_DLC	float	Qualcomm® SA8650P	36.693 ms	1 - 701 MB	NPU
OpenAI-Clip	QNN_DLC	float	Qualcomm® SA8255P	36.693 ms	1 - 701 MB	NPU
OpenAI-Clip	QNN_DLC	float	Qualcomm® QCS8450	46.455 ms	0 - 854 MB	NPU
OpenAI-Clip	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	36.921 ms	1 - 3 MB	NPU
OpenAI-Clip	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	33.277 ms	1 - 1 MB	NPU
OpenAI-Clip	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	17.214 ms	1 - 694 MB	NPU
OpenAI-Clip	QNN_DLC	float	Qualcomm® SA7255P	98.999 ms	1 - 704 MB	NPU
OpenAI-Clip	QNN_DLC	float	Qualcomm® SA8295P	38.524 ms	1 - 798 MB	NPU
OpenAI-Clip	QNN_DLC	float	Snapdragon® 8 Elite Mobile	17.214 ms	1 - 694 MB	NPU
OpenAI-Clip	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	13.331 ms	0 - 704 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Snapdragon® X2 Elite	14.407 ms	0 - 0 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Snapdragon® X Elite	33.346 ms	1 - 1 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Snapdragon® 8 Gen 3 Mobile	22.836 ms	48 - 798 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Qualcomm® Dragonwing™ IQ-8275	29.984 ms	0 - 3 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Qualcomm® Dragonwing™ IQ-8275	67.718 ms	1 - 604 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Qualcomm® Dragonwing™ QCS8550 (Proxy)	32.935 ms	0 - 542 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Qualcomm® SA8775P	32.423 ms	0 - 603 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Qualcomm® SA8650P	32.423 ms	0 - 603 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Qualcomm® SA8255P	32.423 ms	0 - 603 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Qualcomm® Dragonwing™ IQ-9075	32.652 ms	2 - 4 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Qualcomm® Dragonwing™ IQ-X7181	33.346 ms	1 - 1 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-6690	420.338 ms	1 - 686 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-7790	41.427 ms	0 - 658 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-8750	18.398 ms	0 - 632 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Qualcomm® SA7255P	67.718 ms	1 - 604 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Snapdragon® 8 Elite Mobile	18.398 ms	0 - 632 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	14.483 ms	0 - 678 MB	NPU
OpenAI-Clip	QNN_DLC	w8a16	Snapdragon® 7 Gen 4 Mobile	41.427 ms	0 - 658 MB	NPU
OpenAI-Clip	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	21.716 ms	0 - 898 MB	NPU
OpenAI-Clip	TFLITE	float	Snapdragon® 8 Gen 1 Mobile	40.386 ms	0 - 907 MB	NPU
OpenAI-Clip	TFLITE	float	Qualcomm® Dragonwing™ IQ-8275	34.917 ms	0 - 297 MB	NPU
OpenAI-Clip	TFLITE	float	Qualcomm® Dragonwing™ IQ-8275	95.588 ms	0 - 664 MB	NPU
OpenAI-Clip	TFLITE	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	31.241 ms	0 - 3 MB	NPU
OpenAI-Clip	TFLITE	float	Qualcomm® SA8775P	35.93 ms	0 - 670 MB	NPU
OpenAI-Clip	TFLITE	float	Qualcomm® SA8650P	35.93 ms	0 - 670 MB	NPU
OpenAI-Clip	TFLITE	float	Qualcomm® SA8255P	35.93 ms	0 - 670 MB	NPU
OpenAI-Clip	TFLITE	float	Qualcomm® QCS8450	40.386 ms	0 - 907 MB	NPU
OpenAI-Clip	TFLITE	float	Qualcomm® Dragonwing™ IQ-9075	36.095 ms	0 - 296 MB	NPU
OpenAI-Clip	TFLITE	float	Qualcomm® Dragonwing™ Q-8750	17.396 ms	0 - 694 MB	NPU
OpenAI-Clip	TFLITE	float	Qualcomm® SA7255P	95.588 ms	0 - 664 MB	NPU
OpenAI-Clip	TFLITE	float	Qualcomm® SA8295P	38.381 ms	0 - 622 MB	NPU
OpenAI-Clip	TFLITE	float	Snapdragon® 8 Elite Mobile	17.396 ms	0 - 694 MB	NPU
OpenAI-Clip	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	13.007 ms	0 - 538 MB	NPU

License

• The license for the original implementation of OpenAI-Clip can be found here.

References

• Learning Transferable Visual Models From Natural Language Supervision
• 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.