Overview - Qualcomm Dragonwing Documentation

This guide covers AI application development using the tools, runtimes, and frameworks supported on Qualcomm Dragonwing IoT platforms. It is intended for developers who want to train or fine-tune models, prepare them for deployment, and build AI applications on Qualcomm hardware. Most decisions in this guide come down to three questions:

What are you building? Classic ML/vision, generative AI (LLM/VLM), or voice/speech AI — each has its own tools and quick-start paths.
How will you prepare your model? Convert, quantize, compile, or fine-tune it for the target hardware.
How will you run inference? The runtime, container, or integration path that executes the model on the device.

You can bring pretrained models from ONNX, PyTorch, TensorFlow, or LiteRT — or start from Qualcomm’s ready-to-run LLM, VLM, ASR, and TTS models — and run them efficiently across the Qualcomm Kryo™ CPU, Adreno™ GPU, and Hexagon™ NPU (HTP).

This document assume that the device setup has already been completed. If not, please follow the instructions in the Device Setup and Software Update section of EVK before proceeding.

Choose your journey

Use the flowchart below to find the path that fits your application. It walks you through whether you already have a model, what you’re building, how to prepare it, and how to run inference, then links you to the right Ubuntu workflow page. Every highlighted box is clickable.

On-device generative AI availability depends on your Dragonwing development board and model. Start with GenieX, an on device Gen AI inference runtime which either uses llama.cpp runtime (GGML kernels for CPU / GPU / Hexagon HTP) or the Qualcomm® AI Engine Direct runtime (NPU-only). For voice/speech AI, see the Voice AI workflow and Whisper.

Application development & execution flow summary

Flow	Purpose
Edge Impulse	Build and train AI models using audio, image, and other sensor data, or bring your own model in a variety of formats.
Qualcomm IM SDK	Build high-performance camera and video AI pipelines using GStreamer, zero-copy buffers, and GPU pre/post-processing.
GenieX	Run frontier language and vision-language models locally on Hexagon NPU, Adreno GPU, or CPU with a few lines of code.
Qualcomm AI Hub	Download preoptimized models, or bring your own model (BYOM) and optimize, validate, and deploy on hosted Qualcomm platform devices within minutes.
Qualcomm AI Runtime SDK (QAIRT)	Convert, quantize, and compile models from TensorFlow, PyTorch, LiteRT, or ONNX locally for deployment on the NPU.
AIMET (PTQ / QAT)	Recover accuracy lost during quantization using post-training quantization or quantization-aware training.
LiteRT / TFLite	Run `.tflite` models on the NPU (Python or C++) using AI Engine Direct delegates. Works with models from AI Hub, QAIRT, or Edge Impulse.
ONNX Runtime	Run ONNX models on the NPU using ONNX Runtime with AI Engine Direct.
Run models with SNPE or QNN	Run converted `.dlc` or `.so` files on CPU, GPU, or HTP using the QAIRT SDK C/C++ APIs.
Llama.cpp	Run LLMs and VLMs locally using a C++ backend optimized for CPU/GPU and quantized formats.
LLM/VLM using Qualcomm OpenAI container	Run LLMs/VLMs behind a drop-in OpenAI-compatible API on the NPU.
Whisper	Speech transcription, translation, and language identification using the NPU (VoiceAI ASR) or CPU (whisper.cpp).
Audio Analytics container	Speech recognition (ASR), text-to-speech (TTS), and language translation on the NPU.

Prepare your model

Before a model runs efficiently on Qualcomm hardware, it is converted to an executable format and, for the Hexagon NPU (HTP), quantized to a supported precision. (LiteRT models are an exception: they run directly through the AI Engine Direct delegate.) Choose the preparation tool that matches your starting point.

Tool	Use it to	Output
Qualcomm AI Hub	Download a preoptimized model, or bring your own model (BYOM) and have QAIRT compile, convert, and quantize it in the cloud for your target chipset, with no local toolchain required.	Ready-to-run LiteRT or Qualcomm AI Engine Direct model
Qualcomm AI Runtime SDK (QAIRT)	The local alternative to AI Hub BYOM: convert, quantize, and compile models from TensorFlow, PyTorch, LiteRT, or ONNX yourself. Integrates AI Engine Direct and the Neural Processing SDK.	Quantized model / compiled context binary
Qualcomm AI Model Efficiency Toolkit	Recover accuracy lost during quantization using post-training quantization (PTQ) and quantization-aware training (QAT).	Higher-accuracy quantized model
Edge Impulse	Build, train, or fine-tune models from your own audio, image, and sensor data.	Trained model in your chosen format

Run inference

After your model is prepared, choose how to execute it on the device. The runtime you pick depends on your language, model format, and whether you are building a full camera/video pipeline.

Runtime / integration	Best for
LiteRT	High-performance on-device inference from Python or C++, using Qualcomm AI Engine Direct delegates.
QAIRT SDK C++ APIs	Low-level C++ control over model execution and the inference backend.
Qualcomm IM SDK	High-performance camera, video, and vision pipelines that combine capture, preprocessing, inference, and rendering, with zero-copy buffers and GPU pre/post-processing.
Qualcomm® GenAI Inference Engine (Genie)	Running and orchestrating on-device generative AI (LLMs, multimodal) workflows.
Geniex	Fastest way to run LLM/VLM text, image, and audio inference on the NPU via the GenieX CLI.
Llama.cpp	Run LLMs and VLMs locally using a C++ backend optimized for CPU/GPU and quantized formats.

Building a robotics application? The Qualcomm® Intelligent Robotics (QIR) SDK adds ROS-based modules and hardware-accelerated nodes on top of these runtimes.

AI hardware

Qualcomm platforms include the following hardware accelerators for AI inference:

Qualcomm Kryo™ CPU — High-performance CPU with best-in-class power efficiency.
Qualcomm Adreno™ GPU: Balanced power and performance for AI workloads, accelerated with OpenCL kernels. Also used for model pre- and post-processing (for example, the IM SDK runs resize, color conversion, and overlay on the GPU).
Qualcomm Hexagon™ Tensor Processor (HTP): Also known as NPU/DSP/HMX. Optimized for low-power, high-performance AI inference. For best performance, quantize pretrained models to a supported precision.

Key Documents

​Choose your journey

​Application development & execution flow summary

​Prepare your model

​Run inference

​AI hardware

Choose your journey

Application development & execution flow summary

Prepare your model

Run inference

AI hardware