# GenieX
[GenieX](https://geniex.aihub.qualcomm.com/en/get-started/what-is-geniex) is an optimized on-device generative AI inference runtime for Qualcomm platforms designed to make local AI deployments effortless. It enables developers to run frontier LLMs and VLMs efficiently on the Hexagon NPU, Adreno GPU, or CPU using a simple and intuitive programming interface.
The following image shows the GenieX architecture.
GenieX Architecture
## Select an interface
GenieX provides a unified SDK with five developer-facing interfaces. Select the interface that matches your requirements. The following table provides a description and resources for each interface.
GenieX deployment interfaces
| Interface | Description | Resources |
| --- | --- | --- |
| Command-line interface (CLI) | Execute, benchmark, and serve AI models directly from the terminal, enabling rapid experimentation, validation, and deployment workflows. | [GenieX CLI setup](https://geniex.aihub.qualcomm.com/en/run/cli/install#prerequisites) (prerequisites and setup)
[GenieX CLI tutorial](https://geniex.aihub.qualcomm.com/en/run/cli/quickstart) (tutorial) |
| Python | Integrate model inference capabilities into Python applications and services through a streamlined, developer-friendly API. | [GenieX Python setup](https://geniex.aihub.qualcomm.com/en/run/python/install) (prerequisites and setup)
[GenieX Python tutorial](https://geniex.aihub.qualcomm.com/en/run/python/quickstart) (tutorial) |
| Java/Kotlin | Deliver native on-device AI experiences for Android applications with direct access to GenieX runtime capabilities. | [GenieX Kotlin setup](https://geniex.aihub.qualcomm.com/en/run/android/install) (prerequisites and setup)
[GenieX Kotlin tutorial](https://geniex.aihub.qualcomm.com/en/run/android/quickstart#run-your-first-model) (tutorial) |
| Docker | Deploy GenieX within a containerized environment for reproducible, portable, and scalable development and deployment workflows. | [GenieX Docker setup](https://geniex.aihub.qualcomm.com/en/run/linux/install#prerequisites) (prerequisites, setup, and tutorial) |
| OpenAI-Compatible Server | Expose models through an OpenAI API-compatible endpoint, allowing existing OpenAI-based applications and tooling to run locally with minimal or no code changes. | [Local Inference Server](https://geniex.aihub.qualcomm.com/en/run/cli/local-server) (OpenAI-compatible API) |
## GenieX runtimes
The GenieX runtimes enable developers to balance broad model compatibility with optimized Snapdragon performance through a unified deployment stack. GenieX provides two complementary inference runtimes: llama.cpp and Qualcomm AI Engine Direct (QAIRT).
### llama.cpp
Supports execution of any GGUF-formatted model from Hugging Face running on the Hexagon NPU, Adreno GPU, or CPU using Qualcomm’s GGML Hexagon backend. This runtime offers the broadest model ecosystem and maximum deployment flexibility. See [Run a GGUF model from HuggingFace](https://geniex.aihub.qualcomm.com/en/models/supported#run-a-gguf-model-from-hugging-face) for more information about how to deploy and execute a GGUF model.
### QAIRT
Supports deployment of pre-compiled and chipset-optimized model bundles available from the Qualcomm AI Hub. These models are specifically compiled, quantized, and optimized for the target Snapdragon platform, with execution pinned to the Hexagon NPU for maximum performance, power efficiency, and lowest latency. This is the preferred execution path for models from the Qualcomm AI Hub. See [Run a Qualcomm AI Hub model](https://geniex.aihub.qualcomm.com/en/models/supported#run-a-qualcomm-ai-hub-model) for more information about how to deploy and execute an AI Hub model.
See [GenieX runtimes](https://geniex.aihub.qualcomm.com/en/get-started/platforms#geniex-runtimes) for information about when to pick each runtime.
## FAQs
- What platforms does GenieX support?
>
>
> - For compute, GenieX supports Snapdragon X on Windows ARM64. See [Supported platforms](https://geniex.aihub.qualcomm.com/en/get-started/platforms#snapdragon-platforms) for a complete list of supported chipsets for compute, mobile, and IoT.
- What’s the default compute unit (NPU, GPU, or CPU) for each runtime?
>
>
> - By default, both llama.cpp and QAIRT target the NPU when no compute unit is provided. To leverage llama.cpp’s high-performance NPU + CPU per-tensor scheduling mode, specify `hybrid` as the compute unit explicitly.
- How to run a local model?
>
>
> - For locally available GGUF models or Qualcomm AI Hub bundles, follow the [BYOM](https://geniex.aihub.qualcomm.com/en/run/cli/quickstart#run-a-local-model) guide to learn how to deploy and run the local model using GenieX.
- Where are the CLI and API references?
>
>
> - For CLI, see [CLI References](https://geniex.aihub.qualcomm.com/en/run/cli/reference).
> - For Python, see [Python API References](https://geniex.aihub.qualcomm.com/en/run/python/api-reference).
> - For Kotlin, see [Kotlin API References](https://geniex.aihub.qualcomm.com/en/run/android/api-reference).
- Which precision (quantization) should you choose?
>
>
> - For llama.cpp on Snapdragon NPU pick `Q4_0`. It has the best Hexagon NPU support. Qualcomm AI Hub Models are pre-quantized. See [Precision Supported](https://geniex.aihub.qualcomm.com/en/models/supported#precisions-quantizations-supported) for more information.
Last Published: Oct 23, 2026
[Previous Topic
Generative AI](https://docs.qualcomm.com/bundle/publicresource/80-62010-1/topics/gen-ai-llm.md) [Next Topic
Qualcomm Genie](https://docs.qualcomm.com/bundle/publicresource/80-62010-1/topics/genie.md)