GenieX is an on-device inference runtime that runs frontier large language models and vision-language models locally on Qualcomm Snapdragon devices.
GenieX addresses the challenge of running capable AI models on resource-constrained mobile and edge devices without cloud dependencies. It accepts GGUF models from Hugging Face or pre-compiled bundles from Qualcomm AI Hub and dispatches inference across the Hexagon NPU, Adreno GPU, or CPU depending on what is available and optimal. The tool exposes a single underlying C SDK through multiple interfaces: a command-line tool, Python bindings that mirror the Hugging Face transformers API, Kotlin and Java for Android, Docker containers, and an OpenAI-compatible server.
GenieX is purpose-built for Qualcomm Snapdragon hardware and runs on Windows ARM64, Android, and Linux ARM64 platforms. It suits developers building on-device AI applications who want to avoid cloud inference costs and latency, or who need models to run in offline or privacy-sensitive environments. The Python interface follows the familiar transformers pattern of from_pretrained() and generate(), lowering the adoption barrier for developers already working with Hugging Face models. The tool supports a range of models including Llama, Gemma, Granite, and Qwen variants in both language and vision-language forms.
The project maintains active development across multiple runtime paths and platform targets. The codebase is written primarily in Rust, providing memory safety guarantees for a system-level tool. Documentation is comprehensive and organized by platform and interface type, with separate quickstart guides for CLI, Python, Android, and server deployments. The project offers remote device access through Qualcomm Device Cloud for developers without local hardware.
