Model components / Inference code

Qualcomm AI Engine Direct

On-device inference runtime for Qualcomm's Snapdragon processors targeting mobile, automotive, IoT, and PC. Provides unified API across Qualcomm's Hexagon NPU, Adreno GPU, and Kryo CPU with automatic hardware dispatch for optimal performance and power. Powers on-device AI in billions of Android phones and Copilot+ PCs. Supports transformer models with quantization down to INT4. The dominant inference runtime for Android on-device AI workloads.

Qualcomm AI Engine Direct SDK (aka Qualcomm AI Runtime / QNN) for on-device inference on Snapdragon NPU/HTP (Hexagon), GPU (Adreno), CPU (Kryo). Distributed as a proprietary SDK (separate Qualcomm download); open source helper wrapper (quic/ai-engine-direct-helper, BSD-3) exists. Integrated as backend for LiteRT and ExecuTorch. Confirmed live June 2026.

Openness

1 high confidence

1.0

core-sdk: proprietary(Qualcomm AI Runtime SDK, separate EULA download from Qualcomm software center)
open-wrapper: quic/ai-engine-direct-helper(BSD-3, OSI, only a convenience API layer)
hardware-lock: Qualcomm Snapdragon NPU/HTP

The QNN runtime/SDK itself is proprietary (Qualcomm download + license); the open BSD-3 helper merely wraps it and requires the proprietary libraries. Headline engine closed; scored closed.

https://github.com/quic/ai-engine-direct-helper recorded 2026-06-04
BSD-3 helper that 'is an extension of the Qualcomm AI Runtime SDK'; underlying QNN SDK is a separate proprietary download required to run inference on NPU/HTP

Adoption

4 low confidence

4.0

QNN is the on-device inference backend for the broad Snapdragon install base and is wired into LiteRT (Google AI Edge) and ExecuTorch as the Qualcomm NPU backend, giving it wide on-device reach. No clean developer/usage count; level 4 reflects Snapdragon-ecosystem distribution (reported), not a measured figure.

https://github.com/quic/ai-engine-direct-helper recorded 2026-06-04
tooling for running models on Snapdragon NPU/HTP across Windows-on-Snapdragon and Linux ARM64

Capability

4 medium confidence

4.0

Leading mobile/edge on-device inference runtime (HTP acceleration, multi-accelerator, AOT compile) but Snapdragon-locked and edge-scale; C4 within this category, below the server frontier engines.

https://github.com/quic/ai-engine-direct-helper recorded 2026-06-04
runs inference on NPU/HTP; multi-accelerator Qualcomm stack

Unchanged since 2026-06-09 (last edited, not re-checked)