GenieX developer preview: run generative AI on Qualcomm chipsets with just a few lines of code
Sign up for Developer monthly newsletter
Join thousands of developers around the globe who receive latest news and updates from our monthly curated newsletter.
Sign upGenieX is an open-source runtime for running Generative AI models on Qualcomm devices with a few lines of code – built by the Nexa AI team, which joined Qualcomm through acquisition earlier this year. This article walks through what’s included in the Developer Preview and how to run a model across Windows, Android, and Linux.
Earlier this year, Qualcomm acquired Nexa AI, and the Nexa team has shipped a product today that brings a significant update to the Qualcomm AI Hub ecosystem. Existing NexaSDK users can continue using it seamlessly.
We're introducing GenieX — the easiest way to run any Generative AI on Qualcomm devices and available now as a developer preview on GitHub
On-device generative AI has crossed a threshold. Language model capabilities that once required cloud-scale infrastructure can now run efficiently on-device. However, the developer experience is still catching up. Running Gen AI models with NPU acceleration takes more effort than expected, even though the payoff is clear: lower latency and higher throughput at a fraction of the power. Bridging that gap to the NPU is important for developers to build the next generation of on-device AI experiences.
GenieX gives developers a seamless path from optimized Gen AI models to on-device execution, in minutes. This runtime is built around a key principle: run and integrate the latest generative AI models with just a few lines of code, across NPU, GPU and CPU and on Windows, Android and Linux. GenieX is open sourced under the BSD-3-Clause license. It is designed to grow with the open-source community — we welcome everyone to contribute and will feature our contributors.
GenieX extends the Qualcomm AI Stack and provides a streamlined path for developers to bring any powerful AI capabilities on-device, enabling privacy, controlled cloud inference costs and low-latency experiences for end users.
Fast generative AI model onboarding on device
GenieX is the fastest path to run any generative AI models on Qualcomm devices:
- Run any Hugging Face GGUF models.
- Run Gen AI models on Qualcomm AI Hub Models, and we've widened the list even more with latest models like Qwen 3.5, Qwen3, Gemma 4, Granite 4, Ministral-3, Phi-4, GPT-OSS etc.
- Build multimodal and agentic workflows easily.
Out of the box experience with the tools you already use
GenieX combines the power of llama.cpp and Qualcomm AI Runtime (QAIRT) so developers get broad model coverage and optimal NPU performance in one stack. Whatever platform you ship to and whatever language you work in, GenieX gives you the plug-and-play, easy path to Qualcomm silicon on NPU, GPU, and CPU across Windows, Android, and Linux with friendly developer tooling:
- Simple, one-click installation across platforms - Windows, Android, and Linux.
- Windows: CLI (Command Line Interface), Python Package, and local server.
- Android: native Maven package for in-app inference.
- Linux: Docker containers, CLI, Python Package, and local server.
- OpenAI-Compatible Local Server APIs for easy integration with external tools.
Come for support, stay for the community
Get support from experts, connect with like-minded developers, and access exclusive virtual events.
Windows
Command line interface, OpenAI-compatible API.
Fetch and run models with one line of code directly in the terminal and integrate with external local agent tools like OpenClaw.
geniex infer ai-hub-models/Qwen2.5-VL-7B-InstructAndroid
Native Maven package for in-app inference. Integrate generative AI into an Android app using the same idioms developers already use for any other Android library.
Or you can get started with a sample demo app.
val paths = ModelManagerWrapper.getPaths("unsloth/Qwen3.5-2B-GGUF")
?: error("Model not downloaded")
val llm = LlmWrapper.builder()
.llmCreateInput(
LlmCreateInput(
model_name = paths.model_name,
model_path = paths.model_path,
config = ModelConfig(nCtx = 4096),
runtime_id = "llama_cpp",
compute_unit = null, // null → NPU on Snapdragon (recommended)
)
)
.build()
.getOrThrow()
val chat = arrayListOf(ChatMessage("user", "What is 15 times 15"))Linux
Docker containers, CLI, and OpenAI-compatible API. Built for Linux ARM64 systems. Fetch and run models with one line of code.
geniex infer google/gemma-4-E2B-it-qat-q4_0-ggufPython
Import it, load a model, and call it from your own code. Python API is designed with the same Hugging Face transformers experience.
from geniex import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(
"unsloth/GPT-OSS-20B-GGUF",
device_map="auto",
)
messages = [{"role": "user", "content": "What is 2+2?"}]
prompt = model.tokenizer.apply_chat_template(
messages, add_generation_prompt=True,
)
print(model.generate(prompt, max_new_tokens=256).text)What “developer preview” means?
This is an early release for developers who want to start building with GenieX ahead of general availability. The runtime is stable for the workflows above. Between now and GA, expect expanded model coverage, additional integrations, broader platform support, and more.
We're shipping early because we want feedback from real builders, and because we want partners and the wider on-device AI community to start integrating it. We welcome all developers to contribute to our repo.
Next steps
We look forward to seeing the amazing on-device AI experiences you build and hearing your feedback.
- Follow and star our 8K-star GitHub: github.com/qualcomm/geniex
- GenieX landing page: https://aihub.qualcomm.com/geniex
- GenieX documentation: https://geniex.aihub.qualcomm.com/
- Explore the latest Generative AI models on Qualcomm AI Hub Models
- Ask questions in Slack: http://aihub.qualcomm.com/community/slack
Qualcomm acquired Nexa AI and this product is brought you by the former Nexa team. Existing NexaSDK users can continue using it seamlessly.
