qualcomm/ai-hub-models

Qualcomm AI Hub Models is a collection of pre-optimized machine learning models designed for deployment on Qualcomm devices.

The collection addresses the challenge of preparing state-of-the-art models for efficient execution on Qualcomm hardware. Models are pre-optimized for performance metrics including latency and memory consumption. The approach provides compiled assets ready for deployment, along with profiling data from real devices. Users can export models from source code, quantize them as needed, compile for specific target runtimes and chipsets, and evaluate performance on physical hardware through Qualcomm AI Hub Workbench, a cloud-hosted service that handles compilation, profiling, and inference evaluation.

Developers should adopt this collection when targeting Qualcomm devices and seeking models with verified on-device performance characteristics. It suits projects requiring deployment on specific Qualcomm chipsets where pre-compiled, tested models accelerate development. The repository includes Python applications that wrap model inference with preprocessing and postprocessing using PyTorch and NumPy, designed as reference implementations rather than optimized for prediction speed. End-to-end demos allow testing models locally via PyTorch or on cloud-hosted Qualcomm devices. The collection supports multiple on-device runtimes, hardware targets, precision levels, and specific chipsets, with performance data published for each model.

The project maintains an active model directory with consistent export and evaluation workflows. Releases include newly compiled assets and updated performance results generated through the export and evaluate commands. The repository provides both a command-line interface and Python API for browsing, downloading, and working with models. Documentation covers quantization, compilation, profiling, and inference workflows through the Workbench platform. Native sample applications are published separately in a companion repository, while Python applications are included directly within the model definitions.