NexaSDK

NexaSDK lets you build the smartest and fastest on-device AI with minimum energy. It is a highly performant local inference framework that runs the latest multimodal AI models locally on NPU, GPU, and CPU - across Android, Windows, and Linux devices with a few lines of code.

NexaSDK supported latest models weeks or months before anyone else — Qwen3-VL, DeepSeek-OCR, Gemma3n (Vision), and more.

⭐ Star this repo to keep up with exciting updates and new releases about latest on-device AI capabilities.

🏆 Recognized Milestones

• Qualcomm featured us 3 times in official blogs.
  • Innovating Multimodal AI on Qualcomm Hexagon NPU.
  • First-ever Day-0 model support on Qualcomm Hexagon NPU for compute and mobile platforms, Auto and IoT.
  • A simple way to bring on-device AI to smartphones with Snapdragon

🚀 Quick Start

Platform	Links
🖥️ CLI	Quick Start ｜ Docs
🐍 Python	Quick Start ｜ Docs
🤖 Android	Quick Start ｜ Docs
🐳 Linux Docker	Quick Start ｜ Docs

---

🖥️ CLI

Download:

Windows	Linux
arm64 (Qualcomm NPU)	arm64
x64	x64

NPU Access Token (required for NPU models):

Note: Our previous token validation service has been deprecated. For any NPU usage, simply set the access token below — no additional registration or validation is needed.

For Windows:

$env:NEXA_TOKEN="key/eyJhY2NvdW50Ijp7ImlkIjoiNDI1Y2JiNWQtNjk1NC00NDYxLWJiOWMtYzhlZjBiY2JlYzA2In0sInByb2R1Y3QiOnsiaWQiOiJkYjI4ZTNmYy1mMjU4LTQ4ZTctYmNkYi0wZmE4YjRkYTJhNWYifSwicG9saWN5Ijp7ImlkIjoiMmYyOWQyMjctNDVkZS00MzQ3LTg0YTItMjUwNTYwMmEzYzMyIiwiZHVyYXRpb24iOjMxMTA0MDAwMH0sInVzZXIiOnsiaWQiOiI3MGE2YzA4NS1jYjc3LTQ3YmEtOWUxNC1lNjFjYTA2ZThmZjUiLCJlbWFpbCI6ImFsYW40QG5leGE0YWkuY29tIn0sImxpY2Vuc2UiOnsiaWQiOiI4OTlhZGQ2NS1lOTI2LTQ2M2ItODllNi0xMjc0NzM3ZjA1MzYiLCJjcmVhdGVkIjoiMjAyNS0wOS0wNlQwMDo1MzozNi4yMDNaIiwiZXhwaXJ5IjoiMjAzNS0xMi0zMVQyMzo1OTo1OS4wMDBaIn19.BXoUHIEzFMuuZbBT7RvsKO9nTi5950C6kHO64blF7XBnfKvZ6ClA8a55tmszI1ZWdngzpNFTzMM5PV5euuzMCA=="

For Linux / Android adb shell:

export NEXA_TOKEN="key/eyJhY2NvdW50Ijp7ImlkIjoiNDI1Y2JiNWQtNjk1NC00NDYxLWJiOWMtYzhlZjBiY2JlYzA2In0sInByb2R1Y3QiOnsiaWQiOiJkYjI4ZTNmYy1mMjU4LTQ4ZTctYmNkYi0wZmE4YjRkYTJhNWYifSwicG9saWN5Ijp7ImlkIjoiMmYyOWQyMjctNDVkZS00MzQ3LTg0YTItMjUwNTYwMmEzYzMyIiwiZHVyYXRpb24iOjMxMTA0MDAwMH0sInVzZXIiOnsiaWQiOiI3MGE2YzA4NS1jYjc3LTQ3YmEtOWUxNC1lNjFjYTA2ZThmZjUiLCJlbWFpbCI6ImFsYW40QG5leGE4YWkuY29tIn0sImxpY2Vuc2UiOnsiaWQiOiI4OTlhZGQ2NS1lOTI2LTQ2M2ItODllNi0xMjc0NzM3ZjA1MzYiLCJjcmVhdGVkIjoiMjAyNS0wOS0wNlQwMDo1MzozNi4yMDNaIiwiZXhwaXJ5IjoiMjAzNS0xMi0zMVQyMzo1OTo1OS4wMDBaIn19.BXoUHIEzFMuuZbBT7RvsKO9nTi5950C6kHO64blF7XBnfKvZ6ClA8a55tmszI1ZWdngzpNFTzMM5PV5euuzMCA=="

Run your first model:

# Chat with Qwen3
nexa infer ggml-org/Qwen3-1.7B-GGUF

# Multimodal: drag images into the CLI
nexa infer NexaAI/Qwen3-VL-4B-Instruct-GGUF

# NPU (Windows arm64 with Snapdragon X Elite)
nexa infer NexaAI/OmniNeural-4B

• Models: LLM, Multimodal, ASR, OCR, Rerank, Object Detection, Image Generation, Embedding
• Formats: GGUF, NEXA
• 📖 CLI Reference Docs

---

🐍 Python SDK

pip install nexaai

from nexaai import LLM, GenerationConfig, ModelConfig, LlmChatMessage

llm = LLM.from_(model="NexaAI/Qwen3-0.6B-GGUF", config=ModelConfig())

conversation = [
    LlmChatMessage(role="user", content="Hello, tell me a joke")
]
prompt = llm.apply_chat_template(conversation)
for token in llm.generate_stream(prompt, GenerationConfig(max_tokens=100)):
    print(token, end="", flush=True)

• Models: LLM, Multimodal, ASR, OCR, Rerank, Object Detection, Image Generation, Embedding
• Formats: GGUF, NEXA
• 📖 Python SDK Docs

---

🤖 Android SDK

Add to your app/AndroidManifest.xml

<application android:extractNativeLibs="true">

Add to your build.gradle.kts:

dependencies {
    implementation("ai.nexa:core:0.0.19")
}

// Initialize SDK
NexaSdk.getInstance().init(this)

// Load and run model
VlmWrapper.builder()
    .vlmCreateInput(VlmCreateInput(
        model_name = "omni-neural",
        model_path = "/data/data/your.app/files/models/OmniNeural-4B/files-1-1.nexa",
        plugin_id = "npu",
        config = ModelConfig()
    ))
    .build()
    .onSuccess { vlm ->
        vlm.generateStreamFlow("Hello!", GenerationConfig()).collect { print(it) }
    }

• Requirements: Android minSdk 27, Qualcomm Snapdragon 8 Gen 4 Chip
• Models: LLM, Multimodal, ASR, OCR, Rerank, Embedding
• NPU Models: Supported Models
• 📖 Android SDK Docs

---

🐳 Linux Docker

docker pull nexa4ai/nexasdk:latest

export NEXA_TOKEN="your_token_here"
docker run --rm -it --privileged \
  -e NEXA_TOKEN \
  nexa4ai/nexasdk:latest infer NexaAI/Granite-4.0-h-350M-NPU

• Requirements: Qualcomm Dragonwing IQ9, ARM64 systems
• Models: LLM, VLM, ASR, CV, Rerank, Embedding
• NPU Models: Supported Models
• 📖 Linux Docker Docs

---

⚙️ Features & Comparisons

🙏 Acknowledgements

We would like to thank the following projects:

• ggml
• mlx-lm
• mlx-vlm
• mlx-audio

📄 License

NexaSDK uses a dual licensing model:

CPU/GPU Components

Licensed under Apache License 2.0.

NPU Components

• Personal Use: Free license key available from Nexa AI Model Hub. Each key activates 1 device for NPU usage.
• Commercial Use: Contact hello@nexa.ai for licensing.

🤝 Contact & Community Support

Want more model support, backend support, device support or other features? We'd love to hear from you!

Feel free to submit an issue on our GitHub repository with your requests, suggestions, or feedback. Your input helps us prioritize what to build next.