Qualcomm AI Hub Apps: Sample Deployment Recipes for Snapdragon NPUs

The gap between a model file and a running on-device demo

A Qualcomm AI Hub model download gives you weights and a compiled artifact. It does not give you a camera pipeline, a decoder, a UI, or the glue that feeds tensors into the right runtime. Qualcomm AI Hub Apps fills that gap with a directory of small, complete applications, one per task, each paired with one or more models from Qualcomm AI Hub. The README describes the repository as a collection of sample apps and tutorials to help deploy machine learning models on Qualcomm devices. The audience is narrow and specific: an engineer who has picked a Snapdragon target, chosen a model, and now needs a working example of how inference is actually invoked on that hardware. If you are evaluating model accuracy or training, this repository has nothing for you. If you are trying to get an object detector to consume a live camera feed on a phone with a Hexagon NPU, it is the shortest path to a reference implementation.

One directory per task, one runtime per directory

The architecture is deliberately flat. The top-level README carries three tables, one each for Android, Windows and Ubuntu, and each row maps a task to a directory, a language and an inference API. Android entries use Java with C++ where needed, and lean on TensorFlow Lite; the ChatApp there uses the Genie SDK and is tagged for LLM and GenAI work. Windows entries are mostly C++ with ONNX and OpenCV, plus two Python apps for Whisper speech-to-text and Stable Diffusion image generation. Ubuntu entries are Python with TensorFlow Lite and, for the vision apps, GStreamer. That table is the real map of the repository. There is no shared runtime abstraction across platforms, so reading the Android image classification app tells you little about the Windows one beyond the general shape. The supported runtime list is TensorFlow Lite, ONNX and the Genie SDK, which the README describes as a generative AI runtime layered on the Qualcomm AI Engine Direct SDK.

Compute units and the chipsets that actually accelerate

The repository states support for CPU, GPU and NPU, with the NPU entry explicitly including the Hexagon HTP. The chipset list is where the practical constraints appear. NPU acceleration is listed for Snapdragon X2 Elite, X Elite, 8 Elite Gen 5, 8 Elite, 8 Gen 3 and 8 Gen 2, and the README then defers to the QAIRT SDK documentation for all other supported Snapdragon chipsets. The precision requirements are stated plainly: FP16 for floating point on Hexagon Architecture v69 or newer, and INT8 or INT16 for integer on all Snapdragon chipsets. A note at the end of that section says some apps will run without NPU acceleration on non-Snapdragon chipsets. Read that sentence carefully before you plan a port. Running is not the same as accelerating, and the fallback path is the CPU or GPU, which changes the latency profile your application was designed around.

Getting an app onto a device

The getting started instructions are three steps and they are intentionally thin. First, search for your desired OS and app in the tables. Second, open that app's README, which the top-level document says contains the build and installation instructions. There is no repository-wide install command, no single configuration file, and no top-level build script described in the material. Deployment targets are Android 11 (API level 30) or newer, Windows 11, and Ubuntu 24.04 or newer, so the host and target assumptions are already baked in. Contributors are pointed to CONTRIBUTING.md for development setup, repository architecture and how to add new apps. That file is the one to read if you intend to modify an app rather than copy from it, because the top-level README does not describe the internal structure of any individual app. Treat the per-app README as the source of truth for commands, since that is where the project places them.

Where the repository stops being the right tool

The most obvious limitation is hardware reach. The NPU path is Qualcomm silicon, and the README's own note concedes that non-Snapdragon chipsets may run some apps without NPU acceleration. If your product ships on a mix of SoC vendors, this repository gives you an example for one of them. The second limitation is scope: these are sample apps, and the README frames them as samples and tutorials, not as a library you link against. There is no package to install and no versioned API surface documented at the top level. The third is fragmentation. Each app carries its own build instructions, its own language, and its own runtime binding, so an Android Java plus TensorFlow Lite app and a Windows C++ plus ONNX app share a purpose but almost no code. The fourth is that two entries in the tables, WhisperKit for Android and GenieX Chat for Windows, are marked with a footnote stating that source is available on GitHub but not included in the CLI release. If your workflow depends on the CLI release, those are not part of it.

How this differs from ONNX Runtime or TFLite alone

ONNX Runtime and TensorFlow Lite are the engines underneath several of these apps, and both are portable across vendors. The difference is what sits on top. Qualcomm AI Hub Apps adds the Qualcomm-specific path: compiled models from Qualcomm AI Hub, the QAIRT stack, the Hexagon HTP backend, and the Genie SDK for generative workloads. A team using ONNX Runtime directly would write its own session setup, execution provider selection and preprocessing, and would get portability in return. A team using this repository gets a working Snapdragon configuration to copy, and gives up the portability that comes with an engine-first approach. The Genie SDK entries make the split sharper still: they are not an ONNX or TFLite deployment at all, and the LLM tutorials under tutorials/llm_on_genie and tutorials/geniex exist precisely because that path needs its own export and deployment steps.

Release cadence, licence and what it costs to keep up

The release history shows a steady rhythm: v0.34.0 in early August 2026, v0.35.0 two weeks later, v0.36.0 on 2 September 2026, with the last push to the release branch on 10 September 2026. That cadence is the maintenance cost. Apps track Qualcomm AI Hub models and the underlying SDKs, so a model update or an SDK bump can change the build steps in a per-app README you have already copied. Because there is no shared library version to pin, upgrades are per app and manual. The licence is BSD-3-Clause, which the README points to in the LICENSE file. That is a permissive licence, but it governs the sample code, not the models or the SDKs the apps depend on. Those carry their own terms, and the repository does not restate them. If you plan to ship derived code, read the licence files that come with the models and SDKs separately. This is a description of what the repository states, not legal advice.

Editorial conclusion

Adopt Qualcomm AI Hub Apps if you are shipping an Android, Windows 11 or Ubuntu 24.04 application to a Snapdragon device and want a working reference for TFLite, ONNX or Genie SDK inference before writing your own pipeline. Do not adopt it as a general inference framework for x86 Linux servers, iOS, or non-Snapdragon Android hardware, where the NPU path does not apply. Before you commit, verify two things in the README of the specific app directory you intend to copy: the exact model it expects from Qualcomm AI Hub, and the build and installation steps it lists, because those live per app rather than in the top-level document.

Official sources

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