qualcomm/ai-hub-apps
The Qualcomm® AI Hub apps are a collection of state-of-the-art machine learning models optimized for performance (latency, memory etc.) and ready to deploy on Qualcomm® devices.
What it is
The repository is a collection of sample applications and step‑by‑step tutorials that demonstrate how to take a model from the Qualcomm AI Hub and run it on‑device using Qualcomm’s runtimes (TensorFlow Lite, ONNX, or the Genie SDK). Each app is a small, self‑contained project (Android, Windows, or Ubuntu) that you can clone, build, and execute on a supported Snapdragon chipset.
Why it matters
- Shows real‑world deployment – instead of just code snippets, you get full apps (e.g., a chat UI, an object‑detection camera feed, Whisper speech‑to‑text) that already handle model loading, hardware acceleration, and UI integration.
- Covers the whole stack – from the high‑level Android/Kotlin or Windows C++ front‑end down to the low‑level NPU/GPU acceleration via Hexagon HTP or the Genie generative‑AI runtime.
- Accelerates development – developers can copy an app, swap in a different AI Hub model, and immediately see how performance changes across CPU, GPU, and NPU.
Key components
| Platform | Sample tasks | Primary language | Inference API | Notable extras |
|---|---|---|---|---|
| Android | Chat (LLM), Image Classification, Object Detection, Semantic Segmentation, Super‑Resolution, Whisper speech‑to‑text, GenieX chat | Java / Kotlin / C++ | Genie SDK or TensorFlow Lite | OpenCV for camera, live feed support |
| Windows | Same set of tasks as Android (Chat, Classification, Detection, Super‑Resolution, Whisper, Stable Diffusion) | C++ or Python or Go | Genie SDK, ONNX | OpenCV, Python wrappers for Whisper & Stable Diffusion |
| Ubuntu | Hand‑gesture recognition, Pose estimation | Python | TensorFlow Lite | GStreamer pipelines |
How to get started
- Pick your OS and task from the tables in the README.
- Open the sub‑folder’s own README – it contains the exact build commands (Gradle for Android, CMake for Windows C++, pip/venv for Ubuntu Python).
- Install the required Qualcomm SDKs (AI Engine Direct / Genie SDK) and any runtime dependencies (TensorFlow Lite, ONNX Runtime, OpenCV, GStreamer).
- Build and run the app on a device that meets the listed chipset requirements (any Snapdragon with Hexagon HTP for NPU acceleration, otherwise CPU/GPU fallback works).
Supported hardware & runtimes
- Chipsets – All Snapdragon devices that support the QAIRT SDK, explicitly listed from Snapdragon X2 Elite up to Snapdragon 8 Gen 5.
- Compute units – CPU, GPU, and NPU (Hexagon HTP). NPU acceleration requires FP16 or INT8/INT16 weights as noted.
- Operating systems – Android 11 (API 30+) “Red Velvet Cake” and newer, Windows 11, Ubuntu 24.04.
- Runtimes – TensorFlow Lite, ONNX Runtime, and the Genie SDK (a generative‑AI runtime built on Qualcomm’s AI Engine Direct SDK).
Who should use it
- Mobile/edge developers who need a quick reference for integrating AI models on Snapdragon hardware.
- Hardware‑software teams evaluating performance of different compute units (CPU vs. GPU vs. NPU).
- Researchers who want to prototype generative‑AI (LLM chat, Stable Diffusion) on‑device without writing the low‑level glue code.
License
BSD‑3‑Clause – you can freely use, modify, and redistribute the apps, provided you retain the original copyright notice.
All details are taken directly from the repository’s README; no additional features are inferred.