Segment-Anything-Model-3: Optimized for Qualcomm Devices

SAM3 (Segment Anything with Concepts) extends SAM2 with open-vocabulary segmentation, producing bounding boxes and masks for objects matching a natural-language prompt.

This is based on the implementation of Segment-Anything-Model-3 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

Due to licensing restrictions, we cannot distribute pre-exported model assets for this model. 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

See our repository for Segment-Anything-Model-3 on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.semantic_segmentation

Model Stats:

• Input resolution: 1008x1008
• Model checkpoint: sam3
• Model size (SAM3Backbone) (float): 128 MB
• Model size (SAM3Transformer) (float): 23.7 MB
• Number of parameters (SAM3Backbone): 33.5M
• Number of parameters (SAM3Transformer): 6.22M

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
head	PRECOMPILED_QNN_ONNX	float	Snapdragon® X2 Elite	210.959 ms	136 - 136 MB	NPU
head	PRECOMPILED_QNN_ONNX	float	Snapdragon® X Elite	484.987 ms	668 - 668 MB	NPU
head	PRECOMPILED_QNN_ONNX	float	Snapdragon® 8 Gen 3 Mobile	385.658 ms	155 - 168 MB	NPU
head	PRECOMPILED_QNN_ONNX	float	Qualcomm® Dragonwing™ IQ-8275	490.597 ms	106 - 216 MB	NPU
head	PRECOMPILED_QNN_ONNX	float	Qualcomm® Dragonwing™ IQ-9075	501.834 ms	106 - 215 MB	NPU
head	PRECOMPILED_QNN_ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	484.987 ms	668 - 668 MB	NPU
head	PRECOMPILED_QNN_ONNX	float	Qualcomm® Dragonwing™ Q-8750	279.413 ms	131 - 142 MB	NPU
head	PRECOMPILED_QNN_ONNX	float	Snapdragon® 8 Elite Mobile	279.413 ms	131 - 142 MB	NPU
head	PRECOMPILED_QNN_ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	237.002 ms	131 - 143 MB	NPU
head	QNN_CONTEXT_BINARY	float	Snapdragon® X2 Elite	208.783 ms	107 - 107 MB	NPU
head	QNN_CONTEXT_BINARY	float	Snapdragon® X Elite	483.818 ms	107 - 107 MB	NPU
head	QNN_CONTEXT_BINARY	float	Snapdragon® 8 Gen 3 Mobile	385.72 ms	81 - 89 MB	NPU
head	QNN_CONTEXT_BINARY	float	Qualcomm® Dragonwing™ IQ-8275	491.092 ms	106 - 278 MB	NPU
head	QNN_CONTEXT_BINARY	float	Qualcomm® Dragonwing™ IQ-8275	1014.005 ms	91 - 100 MB	NPU
head	QNN_CONTEXT_BINARY	float	Qualcomm® SA8775P	517.449 ms	106 - 117 MB	NPU
head	QNN_CONTEXT_BINARY	float	Qualcomm® SA8650P	517.449 ms	106 - 117 MB	NPU
head	QNN_CONTEXT_BINARY	float	Qualcomm® SA8255P	517.449 ms	106 - 117 MB	NPU
head	QNN_CONTEXT_BINARY	float	Qualcomm® Dragonwing™ IQ-9075	503.102 ms	106 - 278 MB	NPU
head	QNN_CONTEXT_BINARY	float	Qualcomm® Dragonwing™ IQ-X7181	483.818 ms	107 - 107 MB	NPU
head	QNN_CONTEXT_BINARY	float	Qualcomm® Dragonwing™ Q-8750	275.848 ms	61 - 74 MB	NPU
head	QNN_CONTEXT_BINARY	float	Qualcomm® SA7255P	1014.005 ms	91 - 100 MB	NPU
head	QNN_CONTEXT_BINARY	float	Qualcomm® SA8295P	596.76 ms	107 - 112 MB	NPU
head	QNN_CONTEXT_BINARY	float	Snapdragon® 8 Elite Mobile	275.848 ms	61 - 74 MB	NPU
head	QNN_CONTEXT_BINARY	float	Snapdragon® 8 Elite Gen 5 Mobile	238.039 ms	81 - 90 MB	NPU
vision_backbone	PRECOMPILED_QNN_ONNX	float	Snapdragon® X2 Elite	1087.802 ms	265 - 265 MB	NPU
vision_backbone	PRECOMPILED_QNN_ONNX	float	Snapdragon® X Elite	2193.085 ms	943 - 943 MB	NPU
vision_backbone	PRECOMPILED_QNN_ONNX	float	Snapdragon® 8 Gen 3 Mobile	1758.676 ms	20 - 30 MB	NPU
vision_backbone	PRECOMPILED_QNN_ONNX	float	Qualcomm® Dragonwing™ IQ-8275	2528.672 ms	57 - 72 MB	NPU
vision_backbone	PRECOMPILED_QNN_ONNX	float	Qualcomm® Dragonwing™ IQ-9075	2455.563 ms	80 - 94 MB	NPU
vision_backbone	PRECOMPILED_QNN_ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	2193.085 ms	943 - 943 MB	NPU
vision_backbone	PRECOMPILED_QNN_ONNX	float	Qualcomm® Dragonwing™ Q-8750	1409.253 ms	48 - 60 MB	NPU
vision_backbone	PRECOMPILED_QNN_ONNX	float	Snapdragon® 8 Elite Mobile	1409.253 ms	48 - 60 MB	NPU
vision_backbone	PRECOMPILED_QNN_ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	1205.643 ms	8 - 20 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Snapdragon® X2 Elite	1085.756 ms	12 - 12 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Snapdragon® X Elite	2188.961 ms	12 - 12 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Snapdragon® 8 Gen 3 Mobile	1730.548 ms	20 - 27 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Qualcomm® Dragonwing™ IQ-8275	2530.104 ms	12 - 132 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Qualcomm® Dragonwing™ IQ-8275	6363.024 ms	10 - 21 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Qualcomm® SA8775P	2521.158 ms	1 - 11 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Qualcomm® SA8650P	2521.158 ms	1 - 11 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Qualcomm® SA8255P	2521.158 ms	1 - 11 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Qualcomm® Dragonwing™ IQ-9075	2454.801 ms	12 - 131 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Qualcomm® Dragonwing™ IQ-X7181	2188.961 ms	12 - 12 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Qualcomm® Dragonwing™ Q-8750	1412.467 ms	0 - 12 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Qualcomm® SA7255P	6363.024 ms	10 - 21 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Qualcomm® SA8295P	2786.191 ms	2 - 12 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Snapdragon® 8 Elite Mobile	1412.467 ms	0 - 12 MB	NPU
vision_backbone	QNN_CONTEXT_BINARY	float	Snapdragon® 8 Elite Gen 5 Mobile	1215.559 ms	6 - 15 MB	NPU

License

• The license for the original implementation of Segment-Anything-Model-3 can be found here.

References

• SAM 3: Segment Anything with Concepts
• 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.