qualcomm/ai-hub-apps: sample apps for running AI Hub models on Snapdragon NPUs

What qualcomm/ai-hub-apps actually gives you

This is not a framework and not a library you import. It is a directory of sample applications, each paired with one or more models from Qualcomm AI Hub Models, plus tutorials for end-to-end workflows. The README describes the goal plainly: apps optimized for on-device deployment, open-source recipes for running AI Hub Models on local devices, and tutorials.

The audience is narrow and specific. You are building an application that will run on a Snapdragon device and you want a reference implementation for a known task: image classification, object detection, semantic segmentation, super resolution, speech to text, pose estimation, audio classification, portrait segmentation, hand gesture recognition, or a chat app backed by a large language model. The repository gives you the wiring for those tasks on three operating systems.

If you are not on Snapdragon, the value drops sharply. The README notes that some apps will run without NPU acceleration on non-Snapdragon chipsets, but the acceleration story, which is the reason this repository exists, does not transfer.

Three runtimes, three operating systems, one directory per app

The structure is flat and predictable. Each top-level directory is one app, named for the task and the platform: image_classification_android, object_detection_windows_cpp, posenet_ubuntu_py, and so on. The suffix tells you the language and the target. There is no shared build system across the repository, and the README says so indirectly by instructing you to find your app in the tables and then read that app's own README for build and installation instructions.

Inference runs through one of three APIs. TensorFlow Lite covers the Android apps and most of the Ubuntu Python apps. ONNX covers the Windows C++ and Python apps. Genie SDK, described as the generative AI runtime on top of the Qualcomm AI Engine Direct SDK, covers the chat applications: chatapp_android in Java and C++, and chatapp_windows_cpp in C++. One Android entry, geniex_chat_android, uses GenieX in Kotlin and Java.

Supported targets are Android 11 (API v30) and newer, Windows 11, and Ubuntu 24.04 or newer. Compute units are CPU, GPU and NPU, where NPU includes the Hexagon HTP. For NPU acceleration the README states the weight and activation types: FP16 for Snapdragon chipsets with Hexagon Architecture v69 or newer, and INT8 or INT16 for all Snapdragon chipsets. That constraint matters more than the app list, because it decides whether your model can use the NPU at all.

Finding the app and reading its build instructions

There is no repository-wide install command. The README's getting-started section is two steps: find your OS and app in the tables, then follow that app's README. So the first real action is locating the directory for your task and opening the document inside it.

The README gives the clone URL directly, so the starting point is the repository itself:

git clone https://github.com/qualcomm/ai-hub-apps.git

After cloning you have the top-level directories listed in the README, among them the four Ubuntu Python apps: mediapipe_hand_gesture_ubuntu_py, posenet_ubuntu_py, yamnet_ubuntu_py and portrait_segmentation_ubuntu_py. The README's own instruction for what to do next is the second step above: open the selected app and read its README.

For an Android app such as image_classification_android or object_detection_android, the language column says Java and the inference API column says TensorFlow Lite. For a Windows app such as image_classification_windows_cpp or whisper_windows_py, the inference API column says ONNX. Those two columns are the only per-app metadata the top-level README provides; the build steps themselves are inside each app directory.

The top-level README does not describe a rollback or uninstall procedure, and it does not list the commands that build or run any individual app. Treat the app README as the authority on dependencies, the model it expects from Qualcomm AI Hub, and the command that launches it.

Where the repository stops short

The top-level README is an index, and it behaves like one. It does not document model export, conversion, quantization, or how a model gets from Qualcomm AI Hub onto the device. Those steps belong to the model's own documentation and to the individual app READMEs. If you arrive expecting a single end-to-end guide, you will be reading several documents.

Two entries in the tables are marked with an asterisk and the note that the source is available on GitHub but not included in the CLI release: WhisperKit for Android and GenieX Chat for Windows. If you are working from the CLI release rather than a clone, those two are not part of what you get.

Platform coverage is uneven by design. Chat with Genie SDK exists on Android and Windows only. Stable Diffusion image generation exists on Windows only. The Ubuntu set is four Python apps covering gesture, pose, audio and portrait segmentation. There is no iOS target anywhere in the tables, and no macOS target. If your product ships on iPhone, this repository has nothing for you.

The chipset list is also a gate. The README names Snapdragon X2 Elite, X Elite, 8 Elite Gen 5, 8 Elite, 8 Gen 3 and 8 Gen 2, then points to the QAIRT SDK documentation for the full set of supported devices. A device that is not on that list is a device whose NPU path you cannot assume.

How this differs from a general on-device runtime

The obvious alternative is taking a runtime such as TensorFlow Lite or ONNX Runtime on its own and writing the integration yourself. That is a real option, and for a simple model it may be less work than adapting a sample app. The difference in approach is what you inherit. With a bare runtime you own the model conversion, the delegate or execution provider configuration, the preprocessing, and the camera or audio plumbing. With these apps you inherit a working combination of those pieces for a specific task on a specific platform, plus the model pairing from Qualcomm AI Hub.

The trade-off is that a sample app is opinionated about its task. If your pipeline diverges from the sample, you are editing code you did not write rather than composing your own. The other alternative is the tutorials directory, which covers exporting and deploying an LLM with Genie SDK and running and exporting LLMs with the GenieX runtime. Those are workflow documents rather than applications, and they suit a team that wants to understand the deployment path before adopting any app as a base.

Maintenance, licence and upgrade cost

The repository is not archived, and the last push was on 2026-09-10. Releases are frequent: v0.36.0 on 2026-09-02, v0.35.0 on 2026-08-19, v0.34.0 on 2026-08-03. That cadence suggests the sample apps track the underlying runtimes and models rather than sitting still.

That cadence is also the upgrade cost. Because each app is independent and there is no shared build system, an update can touch one directory or several, and you find out by reading release notes and the changed app READMEs. If you fork an app as the base of a product, you are carrying a fork of a moving target. Pinning to a release tag is the practical way to control that, and the repository's tags are the only versioning signal the README points at.

The licence is BSD-3-Clause, stated in the README and present as a LICENSE file at the top level. That is a permissive licence, but this is not legal advice: the apps pull in third-party runtimes, models and SDKs with their own terms, and the repository's licence does not cover those. Check the terms of the Genie SDK, the AI Engine Direct SDK and any model you ship before you rely on the BSD-3 grant alone.

Editorial conclusion

Adopt it if you are shipping to Snapdragon hardware and want a working reference for TensorFlow Lite, ONNX or Genie SDK inference rather than a from-scratch integration. Skip it if your deployment target is not Android 11+, Windows 11 or Ubuntu 24.04+, or if you need an iOS path, which the README does not list. Before committing, open the README of the single app closest to your task and confirm its build steps, the model it expects from Qualcomm AI Hub, and whether your chipset appears in the supported list.

Frequently asked questions

Official sources

1. Issues
2. License: BSD-3-Clause
3. qualcomm/ai-hub-apps on GitHub
4. README
5. Releases