Qualcomm acquires Nexa AI, open-sources GenAI runtime for Hexagon NPUs

GenieX is an on-device Gen AI inference runtime for Qualcomm devices. Bring almost any GGUF model from Hugging Face — or a pre-compiled bundle from Qualcomm AI Hub — and run it locally on the Hexagon NPU, Adreno GPU, or CPU in a few lines of code. One C SDK underneath, exposed through a CLI, Python, Kotlin/Java, Docker, and an OpenAI-compatible server. It is the community version of Qualcomm GENIE.

GenieX runs only on Qualcomm Snapdragon. Find your platform, then jump straight to the interface you want to use.

Platform	Example devices	Jump to a quickstart
🪟 Windows ARM64 (Compute)
Snapdragon X · X Elite

Android*(Mobile)Android SDKLinux ARM64(IoT)*CLI·Docker·PythonNo device on hand? Spin up a remote session on

[Qualcomm Device Cloud].

Pick your interface below. Each one follows the same three steps — Install, Run, and Docs — and shows both runtimes: a GGUF model from Hugging Face (llama_cpp

) and a pre-compiled bundle from Qualcomm AI Hub (qairt

, NPU).

Install

Windows ARM64—download the installer, run it, then open a new terminal.** Linux ARM64**— one line, nosudo

:

curl -fsSL https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-geniex/install.sh | sh

Run — chat with any model in one line (drag in an image for VLMs):

geniex infer google/gemma-4-E4B-it-qat-q4_0-gguf

geniex infer ai-hub-models/Qwen2.5-VL-7B-Instruct

📖 Docs — Install · Quickstart · Command reference

Install

pip install geniex

Run — mirrors Hugging Face transformers

(from_pretrained()

→ .generate()

):

from geniex import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("unsloth/Qwen3.5-2B-GGUF", precision="Q4_0")

messages = [{"role": "user", "content": "What is 2+2?"}]
prompt = model.tokenizer.apply_chat_template(messages, add_generation_prompt=True)

for chunk in model.generate(prompt, max_new_tokens=256, stream=True):
    print(chunk, end="", flush=True)

model.close()
from geniex import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("ai-hub-models/Qwen3-4B")

messages = [{"role": "user", "content": "What is 2+2?"}]
prompt = model.tokenizer.apply_chat_template(messages, add_generation_prompt=True)

for chunk in model.generate(prompt, max_new_tokens=256, stream=True):
    print(chunk, end="", flush=True)

model.close()

📖 Docs — Install · Quickstart · API reference

Install — ships with the CLI (install above).

Run — pull any model (GGUF or Qualcomm AI Hub bundle), then serve an OpenAI-compatible API:

geniex pull ai-hub-models/Qwen3-4B-Instruct-2507
geniex serve   # serves http://127.0.0.1:18181/v1
curl http://127.0.0.1:18181/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ai-hub-models/Qwen3-4B-Instruct-2507",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Point any OpenAI client at http://127.0.0.1:18181/v1

— no code changes.

📖 Docs — Local server guide

Install — add the SDK to your app module's build.gradle.kts

:

dependencies {
    implementation("com.qualcomm.qti:geniex-android:0.3.1")
}

Run — fastest path is the sample app (chat UI, model picker for GGUF + Qualcomm AI Hub bundles, VLM support):

The Android demo app lives in qualcomm/ai-hub-apps. Clone it, open the sample app in Android Studio, and hit

Run.

📖 Docs — Install · Quickstart · API reference

Install

docker pull docker.io/qualcomm/geniex:latest

Run — the container wraps the CLI, so geniex infer …

works exactly as above.

📖 Docs — Docker guide

Install — link against the single C header sdk/include/geniex.h; every other interface is a thin wrapper over it.

📖 Docs — sdk/README.md · notes/build.md

GenieX has two runtimes so you get broad model coverage and peak Snapdragon performance in one stack. Both LLMs and VLMs are supported.

llama.cpp (llama_cpp ) | Qualcomm AI Engine Direct (qairt ) | | |---|---|---| Get models from | |

Qualcomm AI Hub(pre-compiled)FormatCompute unitsBest for For llama.cpp, pick the

precision when prompted — it has the best Hexagon NPU support. See theQ4_0

[Models guide →]for the full list, precisions, and how to run a local model.

Contributions are welcome! Before opening a PR, please read ** CONTRIBUTING.md** for branch naming, commit / PR title format, pre-commit checks, and the FFI-update rule for public SDK headers.

🏗️ Build the CLI, SDK, or Python bindings | |

Run& select compute units / pull modelsnotes/run.md** Release**— SemVer tags, channels, HTP signingnotes/release.md** All developer docs**docs/README.mdQuestions, ideas, or want to show off what you built? Come say hi.

• 💬 — ask questions and chat with the community in real time.Slack - 🐛 — report a bug or request a feature.GitHub Issues - 🔗 — follow Qualcomm AI Hub for news and updates.LinkedIn

Thanks to everyone building GenieX 💙

BSD 3-Clause — see LICENSE and NOTICE.

Use of this project is also subject to Qualcomm's Terms of Use.

source & further reading

github.com — original article