# Overview

## Migrate to LiteRT (Recommended)

TensorFlow Lite is now [LiteRT (Lite Runtime)](https://github.com/google-ai-edge/LiteRT). LiteRT is the successor to TensorFlow Lite and a high-performance runtime for on-device AI. Google states that all future feature updates and performance enhancements will be exclusive to LiteRT. For Qualcomm NPU deployments, new projects should use LiteRT, and users with existing QNN TFLite Delegate workflows should migrate to LiteRT to stay aligned with Google’s future on-device AI runtime direction and access future feature and performance improvements.

LiteRT extends the TensorFlow Lite runtime path with additional deployment capabilities for modern on-device AI workloads. Alongside JIT-based on-device preparation, LiteRT supports an AOT (Ahead-of-Time, offline preparation) flow that can pre-compile models on a development host and deploy optimized artifacts to the target device. As on-device AI workloads increasingly include LLM and GenAI models, these capabilities provide a more scalable deployment path for large models on Qualcomm NPUs.

| Feature | LiteRT | QNN TFLite Delegate |
| --- | --- | --- |
| Compilation flow | AOT + JIT | JIT |
| LLM / GenAI | First-class via LiteRT-LM (Gemma, FastVLM, …) | Limited |
| Graph optimization | GQA→SHA, MHA→SHA, and more | Not available |
| Debug tools | E2E/Per-op accuracy debugger, E2E/Per-op performance profiler | E2E performance profiler |

To get started with LiteRT on Qualcomm:

1. Qualcomm LiteRT document: [litert/vendors/qualcomm/README.md](https://github.com/google-ai-edge/LiteRT/blob/main/litert/vendors/qualcomm/README.md)
2. Migrating from TensorFlow Lite: [Migration guide](https://developers.google.com/edge/litert/migration)
3. Full API reference: [LiteRT developer site](https://developers.google.com/edge/litert)

* * *

## Contents

The TFLite Delegate provides several ways to support running TFLite models on target devices. This library extends the [QNN SDK](https://docs.qualcomm.com/bundle/publicresource/topics/80-63442-10/general_introduction.html), resulting in nearly identical workflows to prepare and execute a model, as well as for implementing an application which uses the model on the target device.

To see exactly what is different about those workflows, you will need to complete both of these tutorials:

1. [Preparing and executing a model with TFLite Delegate](https://docs.qualcomm.com/doc/80-63442-10/topic/tutorial_preparing_and_executing.html).
2. Buliding an application that uses TFLite Delegate:

    1. [C/C++ (for Linux)](https://docs.qualcomm.com/doc/80-63442-10/topic/tutorial_c_application.html)
    2. [Java (for Android)](https://docs.qualcomm.com/doc/80-63442-10/topic/tutorial_java_application.html)

## Navigating These Docs

| Section | Description |
| --- | --- |
| [Tutorials](https://docs.qualcomm.com/doc/80-63442-10/topic/tutorials.html) | These explain how to use this API. This is where you should start. The rest of these docs primarily explain how each feature works in case you need additional context for your specific use case. |
| [Acceleration Support](https://docs.qualcomm.com/doc/80-63442-10/topic/support.html) | Additional context on what specifically is supported for each backend (e.g., operators supported on GPUs). |
| [Tools](https://docs.qualcomm.com/doc/80-63442-10/topic/tools.html) | Reference docs for CLI tools like `qtld-net-run`. |
| [Custom Operator Support](https://docs.qualcomm.com/doc/80-63442-10/topic/custom_op.html) | Detailed instructions for writing your own operations (e.g., `Abs`) if you want to optimize the implementation or add support for a missing operator. |
| [Frequently Asked Questions](https://docs.qualcomm.com/doc/80-63442-10/topic/faq.html) | Common errors and what to do about them. Useful during troubleshooting. |
| [API](https://docs.qualcomm.com/doc/80-63442-10/topic/api.html) | Reference docs for the C, Java, and External APIs. See the tutorials section for example code using these APIs. |
| [Qualcomm® AI Engine Direct Backend Library](https://docs.qualcomm.com/doc/80-63442-10/topic/qnn_libs.html) | This page explains which library files from the QNN SDK are relevant based on your chosen backend (e.g., `libQnnGpu.so` for GPU backends). |
| [Release Notes](https://docs.qualcomm.com/doc/80-63442-10/topic/release_notes.html) | QNN TFLite Delegate release notes log and details. |

## TFLite Delegate Files

These are the files that you may need as you use the TFLite Delegate:

<QNN_SDK_ROOT>
    ├── bin
    │   ├── aarch64-android
    │   │   └── qtld-net-run
    │   ├── aarch64-oe-linux-gcc9.3
    │   │   └── qtld-net-run
    │   ├── aarch64-oe-linux-gcc11.2
    │   │   └── qtld-net-run
    │   └── aarch64-ubuntu-gcc9.4
    │       └── qtld-net-run
    ├── docs
    ├── examples
    ├── include
    │   └── QNN
    │       └── TFLiteDelegate
    │           └── QnnTFLiteDelegate.h
    ├── lib
    │   ├── aarch64-android
    │   │   └── libQnnTFLiteDelegate.so
    │   ├── aarch64-oe-linux-gcc9.3
    │   │   └── libQnnTFLiteDelegate.so
    │   ├── aarch64-oe-linux-gcc11.2
    │   │   └── libQnnTFLiteDelegate.so
    │   ├── aarch64-ubuntu-gcc9.4
    │   │   └── libQnnTFLiteDelegate.so
    │   └── android
    │       └── qtld-release.aar
    ├── LICENSE.pdf
    ├── QNN_TFLITE_DELEGATE_NOTICE.txt
    ├── QNN_TFLITE_DELEGATE_README.txt
    └── QNN_TFLITE_DELEGATE_ReleaseNotes.txt
    Copy to clipboard

Each folder under &lt;QNN\_SDK\_ROOT&gt; has the following artifacts:

| Path / File | Description |
| --- | --- |
| **docs/QNN/TFLiteDelegate** | A folder of HTML files that contain documentation for the Qualcomm® AI Engine Direct Delegate. |
| **examples/QNN/TFLiteDelegate/Models** | Scripts that can generate model. |
| **include/QNN/TFLiteDelegate/QnnTFLiteDelegate.h** | Qualcomm® AI Engine Direct Delegate header file containing the C interface for the delegate. |
| **QNN\_TFLITE\_DELEGATE\_README.txt** | A text file acting as the starting point for this package. |
| **QNN\_TFLITE\_DELEGATE\_ReleaseNotes.txt** | The release notes for the current package. |
| **bin/&lt;architecture&gt;/qtld-net-run** | An example application used to perform inference with the Qualcomm® AI Engine Direct Delegate. See the [Tools](https://docs.qualcomm.com/doc/80-63442-10/topic/tools.html) page for more information. |
| **lib/&lt;architecture&gt;/libQnnTFLiteDelegate.so** | The Qualcomm® AI Engine Direct Delegate dynamic library. |
| **lib/&lt;architecture&gt;/qtld-release.aar** | The Qualcomm® AI Engine Direct Delegate AAR. |

## Next Steps

Learn how to use QNN TFLite Delegate by following the [Preparing and executing a model with TFLite Delegate](https://docs.qualcomm.com/doc/80-63442-10/topic/tutorial_preparing_and_executing.html) tutorial.

Last Published: Aug 06, 2026

[Previous Topic
TFLite Delegate](https://docs.qualcomm.com/bundle/publicresource/80-63442-10/topics/index_TfLite.md) [Next Topic
Tutorials](https://docs.qualcomm.com/bundle/publicresource/80-63442-10/topics/tutorials.md)