You already know why on‑device AI matters. But do you know how to achieve real-time latency targets on Qualcomm-powered devices? Learn how to compile, optimize and test models for a specific device and runtime, while preserving accuracy and staying within memory constraints. To deploy AI on device at scale, you need a repeatable workflow that automates optimization and continually deploys the most efficient model across a diverse portfolio of Qualcomm SoCs. Those practices directly shape real‑world user experiences, from low‑power handheld devices to higher‑performance edge appliances.

This workshop focuses on on‑device AI as it exists in production today: power‑limited, latency‑sensitive, and deployed across millions of devices and chipsets, rather than a single reference platform.

Does this sound familiar?

• A model performs well in training, but accuracy isn’t met on device once it is compiled for the NPU.
• Optimization techniques exist, but it is hard to reason about their downstream impact.
• A model is ready to be deployed on one chipset, but it needs to be deployed and tested against all chipsets users will run it on.

This workshop is designed for teams who are deploying edge AI and want an end-to-end workflow for selecting a runtime, quantizing their model, determining performance bottlenecks and understanding how to integrate a compiled model within their software application.

You will select a model—either bring your own or start with our collection. Then submit jobs through Python APIs, inspecting performance down to the operator and layer level to understand how compilation and quantization choices affect on‑device behavior. Qualcomm AI Hub products provide a workflow for deploying performance-optimized models on-device:

• Qualcomm AI Hub Models, a curated collection of 300+ pre‑optimized and profiled models that serve as starting points for NPU deployment
• Qualcomm AI Hub Workbench, the platform used to optimize (compile and quantize) and test (profile and inference) models on real Qualcomm devices with various runtimes
• Qualcomm AI Hub Apps, which provide sample applications for executing inference and collecting on‑device performance data

Designed for ML and software engineers who want to ship AI models on Qualcomm-powered edge devices, this technical hands-on session strengthens how teams approach selecting and deploying optimized models on Qualcomm NPUs. Participants will leave with a working application, device‑level performance data, and APIs to obtain optimized assets and integrate them directly into existing development workflows. Come walk through hands on examples: run inference on real edge devices, walk through our gesture recognition model and app and discuss constraints of large language models on device. 

Hardware Requirements
Laptop: A device able to run Python commands via a code editor is required (no phone or iPad)
You will also be asked to create an account for Qualcomm AI Hub Workbench prior to the session (more info to follow)

Note: Separate registration and a $25 fee is required.