V-JEPA-2: Optimized for Qualcomm Devices

V-JEPA 2 (Video Joint-Embedding Predictive Architecture 2) is a self-supervised video encoder from Meta AI that uses a ViT (Vision Transformer) backbone to produce per-token video embeddings. It is used as a frozen backbone for downstream video understanding tasks such as action recognition and video retrieval.

This is based on the implementation of V-JEPA-2 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

For more device-specific assets and performance metrics, visit V-JEPA-2 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 V-JEPA-2 on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.video_classification

Model Stats:

• Input resolution: 256x256
• Model checkpoint: facebook/vjepa2-vitl-fpc64-256
• Model size (float): 1.13 GB
• Model size (w8a16): 304 MB
• Number of frames: 8
• Number of parameters: 304M

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
V-JEPA-2	ONNX	float	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	196.793 ms	8 - 2806 MB	NPU
V-JEPA-2	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	259.387 ms	9 - 2587 MB	NPU
V-JEPA-2	ONNX	float	Snapdragon® X2 Elite	187.712 ms	7 - 7 MB	NPU
V-JEPA-2	ONNX	float	Snapdragon® X Elite	431.083 ms	601 - 601 MB	NPU
V-JEPA-2	ONNX	float	Qualcomm® Dragonwing™ IQ-8275	463.704 ms	6 - 16 MB	NPU
V-JEPA-2	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	431.137 ms	0 - 637 MB	NPU
V-JEPA-2	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	470.815 ms	6 - 15 MB	NPU
V-JEPA-2	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	431.083 ms	601 - 601 MB	NPU
V-JEPA-2	ONNX	float	Qualcomm® Dragonwing™ Q-8750	259.387 ms	9 - 2587 MB	NPU
V-JEPA-2	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	219.491 ms	0 - 2608 MB	NPU
V-JEPA-2	QNN_DLC	float	Snapdragon® X2 Elite	167.039 ms	6 - 6 MB	NPU
V-JEPA-2	QNN_DLC	float	Snapdragon® X Elite	396.705 ms	6 - 6 MB	NPU
V-JEPA-2	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	584.698 ms	6 - 3557 MB	NPU
V-JEPA-2	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-8275	435.924 ms	6 - 19 MB	NPU
V-JEPA-2	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	392.929 ms	6 - 1221 MB	NPU
V-JEPA-2	QNN_DLC	float	Qualcomm® QCS8450	584.698 ms	6 - 3557 MB	NPU
V-JEPA-2	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	431.071 ms	6 - 18 MB	NPU
V-JEPA-2	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	396.705 ms	6 - 6 MB	NPU
V-JEPA-2	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	219.491 ms	0 - 2608 MB	NPU
V-JEPA-2	QNN_DLC	float	Qualcomm® SA8295P	533.823 ms	4 - 2167 MB	NPU
V-JEPA-2	QNN_DLC	float	Snapdragon® 8 Elite Gen 6 Mobile	212.722 ms	6 - 2744 MB	NPU

License

• The license for the original implementation of V-JEPA-2 can be found here.

References

• V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
• 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.