EmbedNeural – NPU-Native Multimodal Search for Mobile, IoT and PC

• 📖 Blog Post
• 🎬 Demo on Qualcomm
• 🎬 Demo on Apple

EmbedNeural is an NPU-native multimodal embedding model designed for on-device image and text search, optimized for Apple and Qualcomm NPUs. It enables efficient, privacy-preserving semantic search directly on mobile, IoT, and PC devices—no cloud required.

✨ Key Features

• NPU-native architecture – Purpose-built for Apple Neural Engine and Qualcomm Hexagon NPU, maximizing hardware efficiency.
• Multimodal search – Supports searching across text, images, and audio with unified embeddings.
• Privacy-preserving – All processing happens locally on-device, ensuring data never leaves the user's hardware.
• Cross-platform support – Runs on mobile (iOS/Android), IoT devices, and PC with Apple Silicon or Qualcomm chips.
• Low latency – Optimized quantization and NPU-friendly operators deliver real-time search performance.

📚 Table of Contents

• Use Cases
• Quickstart

🎯 Use Case Examples

Phone: Meme Library Search

Your embarrassing screenshots, personal photos, and saved designs never touch the cloud. Visual search runs 100% locally.

Laptop: Visual Reference Library

EmbedNeural turns your chaotic image library into an instantly searchable visual database—without compromising speed, privacy, or battery life.

🚀 Quickstart

⚠️ Hardware Requirement: EmbedNeural is optimized for Apple Neural Engine and Qualcomm Hexagon NPU.

Step 1: Pull the Model

nexa pull NexaAI/EmbedNeural

Step 2: Start Nexa Serve

nexa serve

Step 3: Install Dependencies

pip install gradio
pip install -r requirements.txt

Step 4: Launch the Demo

python gradio_ui.py