# LiteRT documentation Use the Lite Runtime (LiteRT) framework to convert, optimize, and run LiteRT models with the help of delegates on the Qualcomm^®^ Linux^®^ development kit. ## LiteRT overview High-level LiteRT overview Provides a high-level overview of the LiteRT framework, architecture, delegates, model conversion and quantization methods, and sample applications. ## Get started with running LiteRT models Prerequisites to run LiteRT models Set up the Qualcomm Linux development kit, upgrade it to the latest available software release, and flash the software image. Run a LiteRT model using the GStreamer-based Qualcomm^®^ Intelligent Multimedia SDK Download the required files and use the gst-ai-classification precompiled sample application to run a LiteRT classification model on the Qualcomm Linux development kit. Run a LiteRT model using the native LiteRT sample application Download the required files and use the label\_image native sample application to run a LiteRT classification model on the Qualcomm Linux development kit. ## LiteRT architecture LiteRT on-device inference overview Learn how LiteRT on-device inference loads a model, which is subsequently parsed and executed by the interpreter using a delegate. Accelerate LiteRT models using delegates Use delegates to speed up models efficiently on the CPU, GPU, and specialized Qualcomm hardware, such as the Qualcomm^®^ Adreno^™^ GPU and the Qualcomm^®^ Hexagon^™^ Tensor Processor. Qualcomm^®^ AI Engine direct delegate interface Include the `QnnTFLiteDelegate.h` header and link the appropriate Qualcomm^®^ Neural Network (QNN) delegate library for application compatibility. ## Deploy a LiteRT model Use a pre-optimized LiteRT model Download and use ready-to-deploy LiteRT models from the open-source community or Qualcomm^®^ AI Hub. Convert a TensorFlow model to a LiteRT model Use Python APIs and the `tflite_convert` command to convert models to the LiteRT format. Create an application and run inference Create an application using LiteRT C++ APIs to load a LiteRT model and run inference. Develop a custom application Use the qtimltflite GStreamer-based plug-in to develop your own application and run LiteRT models. ## Run LiteRT sample applications Prerequisites to run LiteRT sample applications Download and copy models, label files, and a sample image to the device to run the label\_image sample application. Run a LiteRT model using an available delegate Run LiteRT models using delegates, such as XNNPACK and GPU, to benchmark model execution. Run the QNN delegate using an external delegate Use the Qualcomm AI Engine direct API as an external delegate, along with the associated libraries, to run the QNN delegate. ## Build LiteRT Optional: Build LiteRT Recompile LiteRT in specific scenarios such as when you want to change the LiteRT library version. Last Published: Oct 09, 2025 Next Topic LiteRT overview