MiniLM-v2: Optimized for Qualcomm Devices

All-MiniLM-L6-v2 maps sentences to a 384-dimensional dense vector space. Trained on 1B+ sentence pairs, it excels at semantic search, clustering, and sentence similarity tasks while being small enough to run on mobile devices.

This is based on the implementation of MiniLM-v2 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 2.50, ONNX Runtime 1.30.0	Download
QNN_DLC	float	Universal	QAIRT 2.50	Download
TFLITE	float	Universal	QAIRT 2.50	Download

For more device-specific assets and performance metrics, visit MiniLM-v2 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 MiniLM-v2 on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.text_generation

Model Stats:

• Embedding dimension: 384
• Input resolution: 128 tokens
• Model checkpoint: sentence-transformers/all-MiniLM-L6-v2
• Model size (float): 86.7 MB
• Number of parameters: 22.7M

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
MiniLM-v2	ONNX	float	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	0.401 ms	0 - 65 MB	NPU
MiniLM-v2	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	0.504 ms	0 - 65 MB	NPU
MiniLM-v2	ONNX	float	Snapdragon® X2 Elite	0.421 ms	1 - 1 MB	NPU
MiniLM-v2	ONNX	float	Snapdragon® X Elite	1.066 ms	45 - 45 MB	NPU
MiniLM-v2	ONNX	float	Snapdragon® 8 Gen 3 Mobile	0.743 ms	0 - 104 MB	NPU
MiniLM-v2	ONNX	float	Snapdragon® 8 Gen 1 Mobile	1.812 ms	0 - 106 MB	NPU
MiniLM-v2	ONNX	float	Qualcomm® Dragonwing™ IQ-8275	1.293 ms	0 - 4 MB	NPU
MiniLM-v2	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	1.049 ms	0 - 12 MB	NPU
MiniLM-v2	ONNX	float	Qualcomm® QCS8450	1.812 ms	0 - 106 MB	NPU
MiniLM-v2	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	1.303 ms	0 - 3 MB	NPU
MiniLM-v2	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	1.066 ms	45 - 45 MB	NPU
MiniLM-v2	ONNX	float	Qualcomm® Dragonwing™ Q-8750	0.504 ms	0 - 65 MB	NPU
MiniLM-v2	ONNX	float	Snapdragon® 8 Elite Extreme Gen 6 Mobile	0.336 ms	0 - 64 MB	NPU
MiniLM-v2	ONNX	float	Snapdragon® 8 Elite Gen 6 Mobile	0.443 ms	0 - 65 MB	NPU
MiniLM-v2	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	0.407 ms	0 - 62 MB	NPU
MiniLM-v2	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	0.506 ms	0 - 62 MB	NPU
MiniLM-v2	QNN_DLC	float	Snapdragon® X2 Elite	0.604 ms	1 - 1 MB	NPU
MiniLM-v2	QNN_DLC	float	Snapdragon® X Elite	1.218 ms	1 - 1 MB	NPU
MiniLM-v2	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	0.733 ms	0 - 104 MB	NPU
MiniLM-v2	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	1.806 ms	0 - 104 MB	NPU
MiniLM-v2	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-8275	1.259 ms	0 - 3 MB	NPU
MiniLM-v2	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	1.041 ms	0 - 2 MB	NPU
MiniLM-v2	QNN_DLC	float	Qualcomm® SA8775P	1.414 ms	0 - 63 MB	NPU
MiniLM-v2	QNN_DLC	float	Qualcomm® SA8650P	1.414 ms	0 - 63 MB	NPU
MiniLM-v2	QNN_DLC	float	Qualcomm® SA8255P	1.414 ms	0 - 63 MB	NPU
MiniLM-v2	QNN_DLC	float	Qualcomm® QCS8450	1.806 ms	0 - 104 MB	NPU
MiniLM-v2	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	1.3 ms	2 - 4 MB	NPU
MiniLM-v2	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	1.218 ms	1 - 1 MB	NPU
MiniLM-v2	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	0.506 ms	0 - 62 MB	NPU
MiniLM-v2	QNN_DLC	float	Qualcomm® SA8295P	2.077 ms	0 - 60 MB	NPU
MiniLM-v2	QNN_DLC	float	Snapdragon® 8 Elite Extreme Gen 6 Mobile	0.323 ms	0 - 64 MB	NPU
MiniLM-v2	QNN_DLC	float	Snapdragon® 8 Elite Gen 6 Mobile	0.44 ms	0 - 65 MB	NPU
MiniLM-v2	TFLITE	float	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	0.436 ms	0 - 59 MB	NPU
MiniLM-v2	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	0.537 ms	0 - 66 MB	NPU
MiniLM-v2	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	0.772 ms	0 - 107 MB	NPU
MiniLM-v2	TFLITE	float	Snapdragon® 8 Gen 1 Mobile	1.873 ms	0 - 106 MB	NPU
MiniLM-v2	TFLITE	float	Qualcomm® Dragonwing™ IQ-8275	1.313 ms	0 - 47 MB	NPU
MiniLM-v2	TFLITE	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	1.054 ms	0 - 2 MB	NPU
MiniLM-v2	TFLITE	float	Qualcomm® QCS8450	1.873 ms	0 - 106 MB	NPU
MiniLM-v2	TFLITE	float	Qualcomm® Dragonwing™ IQ-9075	1.329 ms	0 - 46 MB	NPU
MiniLM-v2	TFLITE	float	Qualcomm® Dragonwing™ Q-8750	0.537 ms	0 - 66 MB	NPU
MiniLM-v2	TFLITE	float	Qualcomm® SA7255P	3.581 ms	0 - 65 MB	NPU
MiniLM-v2	TFLITE	float	Qualcomm® SA8295P	2.159 ms	0 - 59 MB	NPU
MiniLM-v2	TFLITE	float	Snapdragon® 8 Elite Extreme Gen 6 Mobile	0.348 ms	0 - 69 MB	NPU
MiniLM-v2	TFLITE	float	Snapdragon® 8 Elite Gen 6 Mobile	0.467 ms	0 - 70 MB	NPU

License

• The license for the original implementation of MiniLM-v2 can be found here.

References

• Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks
• 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.