# Get started with running LiteRT models
You can run LiteRT models on the Qualcomm^®^ development kits by setting up the prerequisites and downloading the required files. Use either the precompiled gst-ai-classification or the native label\_image sample application to run a LiteRT classification model.
## Prerequisites to run LiteRT models
Before you get started with running LiteRT models, do the following to set up the infrastructure:
Tab Qualcomm Linux
Tab Ubuntu
1. Set up a Qualcomm development kit. For instructions, see the following:
- QCS6490/QCS5430: [Qualcomm Dragonwing RB3 Gen 2 Development Kit quickstart – Linux](https://docs.qualcomm.com/bundle/publicresource/topics/80-70030-253)
- IQ-9075: [Qualcomm Dragonwing IQ-9075 Evaluation Kit quickstart – Linux](https://docs.qualcomm.com/bundle/publicresource/topics/80-70030-263/)
- QCS8275: [Qualcomm IQ-8 Beta Evaluation Kit Quick Start Guide](https://docs.qualcomm.com/bundle/80-70017-263/resource/80-70017-263_REV_AA_Qualcomm_IQ-8_Beta_Evaluation_KitQuick_Start_Guide.pdf)
- IQ-615: [Qualcomm Dragonwing IQ-615 Beta Evaluation Kit Quick Start Guide](https://docs.qualcomm.com/bundle/80-70020-293/resource/80-70020-293_REV_AA_Qualcomm_Dragonwing_IQ-615_Beta_Evaluation_Kit_Quick_Start_Guide.pdf)
>
>
> Note
>
>
> IQ-615 supports inferencing using the CPU and QNN digital signal processor (DSP) delegates only.
Note
The QCS8275 and IQ-615 quick start guides are available for authorized users only. To upgrade your access, go to [Working with Qualcomm](https://www.qualcomm.com/support/working-with-qualcomm).
2. Connect the Qualcomm development kit to a monitor using HDMI.
3. Upgrade the Qualcomm development kit to the latest available software release. For instructions, see [Download the Platform eSDK](https://docs.qualcomm.com/bundle/publicresource/topics/80-70030-254/how_to.html#download-the-platform-esdk).
4. Flash the image to the device. For instructions, see [Flash software images](https://docs.qualcomm.com/bundle/publicresource/topics/80-70030-254/flash_images.html).
1. Set up a Qualcomm development kit. For instructions, see the following:
- QCS6490: [Qualcomm Dragonwing RB3 Gen 2 Development Kit quickstart - Ubuntu](https://docs.qualcomm.com/bundle/publicresource/topics/80-90441-1)
- IQ-9075: [Qualcomm Dragonwing IQ-9075 Evaluation Kit quickstart – Ubuntu](https://docs.qualcomm.com/bundle/publicresource/topics/80-90441-252/iq9-ubuntu-qsg-landing-page-1.html)
2. Connect the Qualcomm development kit to a monitor using HDMI.
3. Upgrade the Qualcomm development kit to the latest available software release.
4. Install LiteRT and sample applications. For instructions, see [Prerequisites](https://docs.qualcomm.com/bundle/publicresource/topics/80-90441-15/sample-app-evk.html).
## Run a LiteRT model using the GStreamer-based Qualcomm IM SDK
Qualcomm development kits contain precompiled LiteRT sample applications to run sample LiteRT models. The gst-ai-classification sample application uses the Qualcomm IM SDK plug-ins to run a LiteRT classification model on Qualcomm development kits. The sample application achieves hardware acceleration using LiteRT delegates.
The following figure shows the pipeline, which receives a video stream from a camera, does the preprocessing, runs the inference on the AI hardware, and displays the results:
Tab Qualcomm Linux
Tab Ubuntu
**Figure: Workflow to run a LiteRT model using Qualcomm IM SDK in Qualcomm Linux**
Note
Running a LiteRT model using the Qualcomm IM SDK is not supported on IQ-615.
**Figure: Workflow to run a LiteRT model using Qualcomm IM SDK in Ubuntu**
The gst-ai-classification sample application does the following:
1. Opens the IMX577 camera on the Qualcomm development kit with a specific resolution and frame rate; for example, 1080p at 30 fps
2. Preprocesses each camera frame to provide the input data to a classification model
For example, the gst-ai-classification sample application:
1. Downscales a 1080p frame to a 224x224 resolution
2. Normalizes the input frame based on the model requirements
3. The qtimltflite Qualcomm IM SDK plug-in, built on top of the LiteRT C++ API, does the following:
1. Loads the sample LiteRT classification model
2. Performs inference on the model using hardware acceleration
4. Postprocesses the output from the inference, that is, extracts the label with the highest predicted probability within the output tensor
5. Overlays the inference result on the original camera input image and displays it on the connected monitor
### Download a sample model and label file
To run a precompiled LiteRT model with the gst-ai-classification sample application, a LiteRT model and its label file must be available on the device.
To download a sample LiteRT model and the corresponding label file, and then copy them to the device, do the following:
1. Go to [Qualcomm^®^ AI Hub](https://aihub.qualcomm.com/models/inception_v3?searchTerm=ince&chipsets=qualcomm-qcs6490-proxy), and download the Inception-v3 quantized model.
Note
The gst-ai-classification sample application is demonstrated for QCS6490.
2. To download the corresponding label file, run the following command:
curl -L -O https://raw.githubusercontent.com/quic/sample-apps-for-qualcomm-linux/refs/heads/main/artifacts/json_labels/classification.json
Copy to clipboard
Note
- The model is available on Qualcomm AI Hub and the corresponding label file is available on QUIC GitHub.
- Commands use placeholder model names for Qualcomm AI Hub models. Replace each placeholder with the corresponding model name from Qualcomm AI Hub.
3. To copy the models and label files to the device using the secure copy protocol (SCP), run the following commands:
Tab Qualcomm Linux
Tab Ubuntu
# For SCP, run the following command:
ssh root@[ip-addr]
mount -o remount,rw /usr
cd /etc
mkdir labels
mkdir media
exit
Copy to clipboard
# Copy files securely
scp classification.json root@[ip-addr]:/etc/labels
scp inception_v3-inception-v3-w8a8.tflite root@[ip-addr]:/etc/models
Copy to clipboard
1. To get the IP address of a Qualcomm development kit, run the following command:
ifconfig wlan0
Copy to clipboard
2. When prompted for a password, enter `oelinux123`.
# For SCP, run the following command:
ssh ubuntu@[ip-addr]
exit
Copy to clipboard
# Copy files securely
scp classification.json ubuntu@[ip-addr]:/home/ubuntu
scp inception_v3-inception-v3-w8a8.tflite ubuntu@[ip-addr]:/home/ubuntu
ssh ubuntu@[ip-addr]
sudo mkdir /etc/models
sudo mkdir /etc/labels
sudo mkdir /etc/media
sudo cp /home/ubuntu/inception_v3-inception-v3-w8a8.tflite /etc/models
sudo cp /home/ubuntu/classification.json /etc/labels
Copy to clipboard
To get the IP address of a Qualcomm development kit, run the following command:
ifconfig wlan0
Copy to clipboard
### Run a LiteRT model with the gst-ai-classification sample application
After copying a sample model and the corresponding label file to the device, do the following to run the LiteRT model:
1. To run inference using LiteRT, run the following command:
Tab Qualcomm Linux
Tab Ubuntu
ssh root@[ip-addr]
Copy to clipboard
ssh ubuntu@[ip-addr]
Copy to clipboard
1. To set up the Wayland Display environment, run the following command:
Tab Qualcomm Linux
Tab Ubuntu
export XDG_RUNTIME_DIR=/dev/socket/weston && export WAYLAND_DISPLAY=wayland-1
Copy to clipboard
export XDG_RUNTIME_DIR=/run/user/$(id -u ubuntu)&& export WAYLAND_ DISPLAY=wayland-1
Copy to clipboard
2. Modify the `config_classification.json` file in the `/etc/configs` folder, as follows:
Tab Qualcomm Linux
Tab Ubuntu
>
>
> {
> "file-path": "/etc/media/video.mp4",
> "ml-framework": "tflite",
> "model": "/etc/models/inception_v3-inception-v3-w8a8.tflite",
> "labels": "/etc/labels/classification.json",
> "threshold": 40,
> "runtime": "dsp",
> "output-type": "waylandsink"
> }
> Copy to clipboard
1. Push the `video.mp4` file to the `/etc/media` folder.
The default path for the video file is `/etc/media/video.mp4`, labels path is `/etc/labels/classification.json`, and model is `/etc/model/inception_v3-inception-v3-w8a8.tflite`.
>
>
> {
> "file-path": "/etc/media/video.mp4",
> "ml-framework": "tflite",
> "model": "/etc/models/inception_v3-inception-v3-w8a8.tflite",
> "labels": "/etc/labels/classification.json",
> "threshold": 40,
> "runtime": "dsp",
> "output-type": "waylandsink"
> }
> Copy to clipboard
1. Push the `video.mp4` file to the `/etc/media` folder.
The default path for the video file is `/etc/media/video.mp4`, labels path is `/etc/labels/classification.json`, and model is `/etc/model/inception_v3-inception-v3-w8a8.tflite`.
>
>
> Note
>
>
> This step requires sudo access.
3. Run the classification sample application:
gst-ai-classification --config-file=/etc/configs/config_classification.json
Copy to clipboard
2. To run the sample application using a custom classification model and labels file, use the following arguments:
- `--model`
- `--labels`
1. Modify the `config_classification.json` file in the `/etc/configs` folder, as follows:
{
"file-path": "/etc/media/video.mp4",
"model":"/etc/models/custom_model.tflite",
"ml-framework": "tflite",
"labels": "/etc/labels/custom_labels.json",
"output-type": "waylandsink"
}
Copy to clipboard
2. To run the classification sample application, run the following command:
gst-ai-classification --config-file=/etc/configs/config_classification.json
Copy to clipboard
3. To stop the sample application, select **Ctrl+C**.
When the sample application is running, it displays the video stream on the connected monitor with inference results overlaid on the frame.
## Run a LiteRT model using the native LiteRT sample application
You can run LiteRT models using a sample LiteRT application called label\_image, which is a part of the TensorFlow repository. The label\_image sample application and the LiteRT library are cross-compiled and installed on the target device.
The label\_image sample application does the following:
1. Loads a classification LiteRT model
2. Performs inference on an image using a delegate to speed up the model on the Qualcomm hardware
To run a model using the label\_image sample application, do the following:
Tab Qualcomm Linux
Tab Ubuntu
1. Download the sample model, corresponding labels, and an example image:
- [BMP file](https://github.com/sourcecode369/tensorflow-1/tree/master/tensorflow/lite/examples/label_image/testdata/)
- [MobileNet LiteRT model](https://github.com/emgucv/models/blob/master/mobilenet_v1_1.0_224_float_2017_11_08/mobilenet_v1_1.0_224.tflite)
2. Run the following commands on the host computer:
wget http://download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_1.0_224_quant.tgz
Copy to clipboard
tar -xvf mobilenet_v1_1.0_224_quant.tgz
Copy to clipboard
wget https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_1.0_224_frozen.tgz
Copy to clipboard
tar -xvf mobilenet_v1_1.0_224_frozen.tgz
Copy to clipboard
# For SCP, run the following command:
ssh root@[ip-addr]
mount -o remount,rw /usr
cd /etc
mkdir artifacts
exit
Copy to clipboard
scp mobilenet_v1_1.0_224_quant.tflite root@[ip-addr]:/etc/artifacts
scp grace_hopper.bmp root@[ip-addr]:/etc/artifacts
scp mobilenet_v1_1.0_224/labels.txt root@[ip-addr]:/etc/artifacts
scp mobilenet_v1_1.0_224.tflite root@[ip-addr]:/etc/artifacts
Copy to clipboard
3. To run an inference using either of the following delegates, do the following:
>
>
> - To run the model on the Arm^®^ CPU using the XNNPACK delegate:
>
>
>
> label_image -l /etc/artifacts/labels.txt -i /etc/artifacts/grace_hopper.bmp -m /etc/artifacts/mobilenet_v1_1.0_224_quant.tflite -c 10 -p 1 --xnnpack_delegate 1
> Copy to clipboard
>
> - To run the model on the Qualcomm^®^ Adreno^™^ GPU using the GPU delegate:
>
>
>
> label_image -l /etc/artifacts/labels.txt -i /etc/artifacts/grace_hopper.bmp -m /etc/artifacts/mobilenet_v1_1.0_224.tflite -c 10 -p 1 --gl_backend 1
> Copy to clipboard
1. Download the sample model, corresponding labels, and an example image:
- [BMP file](https://github.com/sourcecode369/tensorflow-1/tree/master/tensorflow/lite/examples/label_image/testdata/)
- [MobileNet LiteRT model](https://github.com/emgucv/models/blob/master/mobilenet_v1_1.0_224_float_2017_11_08/mobilenet_v1_1.0_224.tflite)
2. Run the following commands:
sudo mkdir /etc/artifacts
cd /etc/artifacts
Copy to clipboard
wget http://download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_1.0_224_quant.tgz
Copy to clipboard
tar -xvf mobilenet_v1_1.0_224_quant.tgz
Copy to clipboard
wget https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_1.0_224_frozen.tgz
Copy to clipboard
tar -xvf mobilenet_v1_1.0_224_frozen.tgz
Copy to clipboard
3. Copy the image and model files downloaded in Step 1 to `/etc/artifacts`.
4. To run an inference using either of the following delegates, do the following:
- To run the model on the Arm^®^ CPU using the XNNPACK delegate:
>
>
> label_image -l /etc/artifacts/labels.txt -i /etc/artifacts/grace_hopper.bmp -m /etc/artifacts/mobilenet_v1_1.0_224_quant.tflite -c 10 -p 1 --xnnpack_delegate 1
> Copy to clipboard
- To run the model on the Adreno using the GPU delegate:
>
>
> label_image -l /etc/artifacts/labels.txt -i /etc/artifacts/grace_hopper.bmp -m /etc/artifacts/mobilenet_v1_1.0_224.tflite -c 10 -p 1 --gl_backend 1
> Copy to clipboard
## Next steps
- [Deploy a LiteRT model](https://docs.qualcomm.com/doc/80-70030-54/topic/tensorflow-lite-developer-workflow.html#tensorflow-lite-developer-workflow)
- [Run LiteRT sample applications](https://docs.qualcomm.com/doc/80-70030-54/topic/sample-applications.html#run-litert-sample-apps)
Last Published: Jul 09, 2026
[Previous Topic
LiteRT overview](https://docs.qualcomm.com/bundle/publicresource/80-70030-54/topics/tflite-landing-page.md) [Next Topic
LiteRT architecture](https://docs.qualcomm.com/bundle/publicresource/80-70030-54/topics/arch.md)