Phi-4-Mini-Instruct: Optimized for Qualcomm Devices

Phi-4-mini-instruct is a lightweight open model built upon synthetic data and filtered publicly available websites - with a focus on high-quality, reasoning dense data.

This is based on the implementation of Phi-4-Mini-Instruct 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 Phi-4-Mini-Instruct on-device

Follow the GenieX quickstart to install GenieX and deploy the model on a target 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
GENIEX_LLAMACPP	q4_0	Universal		Download

For more device-specific assets and performance metrics, visit Phi-4-Mini-Instruct 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 Phi-4-Mini-Instruct on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.text_generation

Model Stats:

• Number of parameters: 3.8B
• 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. The lower bound is for a short prompt (up to 128 tokens, i.e., one iteration of the prompt processor) and the upper bound is for a prompt using the full context length (4096 tokens).
• Response Rate: Rate of response generation after the first response token.

Performance Summary

Model	Runtime	Precision	Chipset	Context Length	Response Rate (tokens per second)	Time To First Token (range, seconds)
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® 8 Elite Gen 5 Mobile	512	23.272858	0.92645325 - 3.705813
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® 8 Elite Gen 5 Mobile	512	22.664968	1.03434575 - 4.137383
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® 8 Elite Gen 5 Mobile	512	19.242405	0.17659049999999998 - 0.7063619999999999
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® 8 Elite Gen 5 Mobile	4096	12.621046	2.2515993749999996 - 72.05117999999999
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® 8 Elite Gen 5 Mobile	4096	12.03971	2.4794271875 - 79.34167
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® 8 Elite Gen 5 Mobile	4096	11.220493	0.36033615625 - 11.530757
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® 8 Elite Mobile	512	24.055461	0.9468802500000001 - 3.7875210000000004
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® 8 Elite Mobile	512	23.939526	0.95809125 - 3.832365
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® 8 Elite Mobile	512	19.557942	0.1830165 - 0.732066
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® 8 Elite Mobile	4096	13.965758	1.94027484375 - 62.088795
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® 8 Elite Mobile	4096	8.177678	2.3551633124999998 - 75.36522599999999
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® 8 Elite Mobile	4096	13.116693	0.35286768749999997 - 11.291765999999999
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® X2 Elite	512	33.714527	0.24767750000000002 - 0.9907100000000001
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® X2 Elite	512	33.35471	0.248513 - 0.994052
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® X2 Elite	512	22.484254	0.12369325 - 0.494773
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® X2 Elite	4096	24.554353	0.44789846875 - 14.332751
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® X2 Elite	4096	24.309187	0.4479129375 - 14.333214
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® X2 Elite	4096	15.507219	0.19384584375 - 6.203067
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® X Elite	512	27.988389	0.40832025 - 1.633281
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® X Elite	512	27.143556	0.46132475 - 1.845299
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® X Elite	512	16.042448	0.25184725 - 1.007389
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® X Elite	4096	15.617096	0.913791375 - 29.241324
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® X Elite	4096	16.000366	0.9590768749999999 - 30.690459999999998
Phi-4-Mini-Instruct	GENIEX_LLAMACPP	q4_0	Snapdragon® X Elite	4096	9.673267	0.41174053125 - 13.175697

License

• The license for the original implementation of Phi-4-Mini-Instruct can be found here.

References

• Phi-4 Technical Report
• 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