Intern3.5-VL-4B: Optimized for Qualcomm Devices

InternVL3.5 is a vision-language model from OpenGVLab capable of understanding both text and images for multimodal reasoning tasks such as visual question answering and image captioning.

This is based on the implementation of Intern3.5-VL-4B 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.

Deploying Intern3.5-VL-4B on-device

Follow the GenieX quickstart to install GenieX and deploy the model on a target device.

See the LLM-on-Genie tutorial to run with the Genie runtime. Note: Genie support will be deprecated soon.

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
GENIE	w4a16	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	QAIRT 2.45	Download
GENIE	w4a16	Snapdragon® 8 Elite For Galaxy Mobile	QAIRT 2.45	Download
GENIEX_QAIRT	w4a16	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	QAIRT 2.45	Download
GENIEX_QAIRT	w4a16	Snapdragon® 8 Elite For Galaxy Mobile	QAIRT 2.45	Download
GENIEX_QAIRT	w4a16	Snapdragon® X2 Elite	QAIRT 2.45	Download
GENIEX_QAIRT	w4a16	Snapdragon® X Elite	QAIRT 2.45	Download
GENIEX_QAIRT	w4a16	Qualcomm® Dragonwing™ IQ-8275	QAIRT 2.45	Download
GENIEX_QAIRT	w4a16	Qualcomm® Dragonwing™ IQ-9075	QAIRT 2.45	Download

For more device-specific assets and performance metrics, visit Intern3.5-VL-4B 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 Intern3.5-VL-4B on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.text_generation

Model Stats:

• Model architecture: Multimodal Transformer with InternViT Vision Encoder, Qwen-based Large Language Model (LLM) backend, Grouped Query Attention (GQA), and SwiGLU activation. It follows a ViT → MLP → LLM design for efficient vision-language understanding and reasoning.
• Response Rate: Rate of response generation after the first response token.
• Supported languages: No exact mention
• TTFT: Time To First Token is the time it takes to generate the first response token. This is expressed as a range because it varies based on the length of the prompt.

Performance Summary

Model	Runtime	Precision	Chipset	Context Length	Response Rate (tokens per second)	Time To First Token (range, seconds)
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	512	16.16788	0.094894 - 0.379576
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	1024	16.833125	0.089904 - 0.719232
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	2048	16.129407	0.088315 - 1.41304
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile	4096	15.85137	0.093515 - 2.99248
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Snapdragon® 8 Elite For Galaxy Mobile	512	16.368874	0.090156 - 0.360624
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Snapdragon® 8 Elite For Galaxy Mobile	1024	16.195512	0.084252 - 0.674016
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Snapdragon® 8 Elite For Galaxy Mobile	2048	15.670499	0.085424 - 1.366784
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Snapdragon® 8 Elite For Galaxy Mobile	4096	16.044207	0.08146099999999999 - 2.6067519999999997
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Snapdragon® X2 Elite	4096	37.188501	0.050552 - 1.617664
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Snapdragon® X Elite	4096	20.489623	0.099245 - 3.17584
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Qualcomm® Dragonwing™ Q-8750	512	16.368874	0.090156 - 0.360624
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Qualcomm® Dragonwing™ Q-8750	1024	16.195512	0.084252 - 0.674016
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Qualcomm® Dragonwing™ Q-8750	2048	15.670499	0.085424 - 1.366784
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Qualcomm® Dragonwing™ IQ-X7181	4096	20.489623	0.099245 - 3.17584
Intern3.5-VL-4B	GENIEX_QAIRT	w4a16	Qualcomm® Dragonwing™ Q-8750	4096	16.044207	0.08146099999999999 - 2.6067519999999997
Vision-Encoder	GENIEX_QAIRT	w4a16	Snapdragon® 8 Elite Gen 5 For Galaxy Mobile
Vision-Encoder	GENIEX_QAIRT	w4a16	Snapdragon® 8 Elite For Galaxy Mobile
Vision-Encoder	GENIEX_QAIRT	w4a16	Snapdragon® X2 Elite
Vision-Encoder	GENIEX_QAIRT	w4a16	Snapdragon® X Elite
Vision-Encoder	GENIEX_QAIRT	w4a16	Qualcomm® Dragonwing™ IQ-8275
Vision-Encoder	GENIEX_QAIRT	w4a16	Qualcomm® Dragonwing™ IQ-9075
Vision-Encoder	GENIEX_QAIRT	w4a16	Qualcomm® Dragonwing™ IQ-X7181
Vision-Encoder	GENIEX_QAIRT	w4a16	Qualcomm® Dragonwing™ Q-8750

License

• The license for the original implementation of Intern3.5-VL-4B can be found here.

References

• InternVL3.5 Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency
• 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.

Usage and Limitations

This model may not be used for or in connection with any of the following applications:

• Accessing essential private and public services and benefits;
• Administration of justice and democratic processes;
• Assessing or recognizing the emotional state of a person;
• Biometric and biometrics-based systems, including categorization of persons based on sensitive characteristics;
• Education and vocational training;
• Employment and workers management;
• Exploitation of the vulnerabilities of persons resulting in harmful behavior;
• General purpose social scoring;
• Law enforcement;
• Management and operation of critical infrastructure;
• Migration, asylum and border control management;
• Predictive policing;
• Real-time remote biometric identification in public spaces;
• Recommender systems of social media platforms;
• Scraping of facial images (from the internet or otherwise); and/or
• Subliminal manipulation