Gemma-4-E4B-it: Optimized for Qualcomm Devices
Gemma is a family of open models built by Google DeepMind. Gemma 4 models are multimodal, handling text and image input (with audio supported on small models) and generating text output. This release includes open-weights models in both pre-trained and instruction-tuned variants. Gemma 4 features a context window of up to 256K tokens and maintains multilingual support in over 140 languages.
This is based on the implementation of Gemma-4-E4B-it 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 Gemma-4-E4B-it 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 Gemma-4-E4B-it 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 Gemma-4-E4B-it on GitHub for usage instructions.
Model Details
Model Type: Model_use_case.text_generation
Model Stats:
- Model architecture: Mixture-of-Experts (MoE) Transformer with Per-Layer Expert Selection and selective routing.
- Supported languages: Multilingual (trained on 140+ languages)
- 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.
- 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) |
|---|---|---|---|---|---|---|
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Gen 5 Mobile | 512 | 17.905892 | 1.12984925 - 4.519397 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Gen 5 Mobile | 512 | 17.771069 | 1.2465307499999998 - 4.986122999999999 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Gen 5 Mobile | 512 | 16.824563 | 0.195887 - 0.783548 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Gen 5 Mobile | 4096 | 13.347809 | 2.2231766250000002 - 71.14165200000001 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Gen 5 Mobile | 4096 | 11.701253 | 2.24031025 - 71.689928 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Gen 5 Mobile | 4096 | 13.26779 | 0.26545246875 - 8.494479 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Mobile | 512 | 17.234032 | 1.26287975 - 5.051519 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Mobile | 512 | 16.777481 | 1.35680775 - 5.427231 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Mobile | 512 | 16.021655 | 0.1988185 - 0.795274 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Mobile | 4096 | 12.238845 | 2.3695864375 - 75.826766 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Mobile | 4096 | 12.105188 | 2.42485496875 - 77.595359 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® 8 Elite Mobile | 4096 | 12.358648 | 0.29968165625 - 9.589813 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® X2 Elite | 512 | 23.827865 | 0.6804272499999999 - 2.7217089999999997 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® X2 Elite | 512 | 25.136049 | 0.729007 - 2.916028 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® X2 Elite | 512 | 20.846358 | 0.14691300000000002 - 0.5876520000000001 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® X2 Elite | 4096 | 12.013094 | 1.25067746875 - 40.021679 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® X2 Elite | 4096 | 13.238633 | 1.14075040625 - 36.504013 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® X2 Elite | 4096 | 17.911474 | 0.19384078124999998 - 6.2029049999999994 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® X Elite | 512 | 20.525636 | 0.53967775 - 2.158711 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® X Elite | 512 | 19.702914 | 0.6135875 - 2.45435 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® X Elite | 512 | 11.882264 | 0.35202125 - 1.408085 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® X Elite | 4096 | 11.823266 | 0.80318884375 - 25.702043 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® X Elite | 4096 | 8.192524 | 0.93255509375 - 29.841763 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Snapdragon® X Elite | 4096 | 10.034623 | 0.4641780625 - 14.853698 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Qualcomm® QCS9075 | 512 | 13.439978 | 1.7480915 - 6.992366 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Qualcomm® QCS9075 | 512 | 13.431804 | 1.75437625 - 7.017505 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Qualcomm® QCS9075 | 512 | 8.859328 | 0.3919085 - 1.567634 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Qualcomm® QCS9075 | 4096 | 10.071341 | 2.1266885312499997 - 68.05403299999999 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Qualcomm® QCS9075 | 4096 | 10.054127 | 2.0949564375 - 67.038606 |
| Gemma-4-E4B-it | GENIEX_LLAMACPP | q4_0 | Qualcomm® QCS9075 | 4096 | 7.626485 | 0.5007065625 - 16.02261 |
License
- The license for the original implementation of Gemma-4-E4B-it can be found here.
References
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