| # EmbedNeural |
|
|
| *On-device multimodal embedding model enabling instant, private NPU-powered visual search.* |
|
|
| --- |
|
|
| ## Quickstart |
|
|
| [Instruction](https://sdk.nexa.ai/model/EmbedNeural) |
|
|
| ## Model Description |
|
|
| **EmbedNeural** is the world’s first multimodal embedding model purpose-built for **Qualcomm Hexagon NPU** devices. It enables **instant, private, battery-efficient** natural-language image search directly on laptops, phones, XR, and edge devices — with no cloud and no uploads. |
|
|
| The model continuously indexes local images using NPU acceleration, turning unorganized photo folders into a fully searchable visual database that runs entirely on-device. |
|
|
| --- |
|
|
| ## Key Features |
|
|
| ### ⚡ NPU-accelerated multimodal embeddings |
| Optimized for Qualcomm NPUs to deliver sub-second search and dramatically lower power consumption. |
|
|
| ### 🔍 Natural-language visual search |
| Query thousands of images instantly using everyday language (e.g., “green bedroom aesthetic”, “cat wearing sunglasses”). |
|
|
| ### 🔒 100% local and private |
| All computation stays on-device. No cloud. No upload. No tracking. |
|
|
| ### 🔋 Ultra-low power |
| Continuous background indexing uses ~10× less power than CPU/GPU methods, enabling true always-on search. |
|
|
| --- |
|
|
| ## Why It Matters |
|
|
| People save thousands of images — memes, screenshots, design inspo, photos — but struggle to find them when needed. Cloud solutions compromise privacy; CPU/GPU search drains battery. |
|
|
| EmbedNeural removes these tradeoffs by combining: |
| - **Instant retrieval** (~0.03s across thousands of images) |
| - **Continuous local indexing** |
| - **Zero data upload** |
| - **NPU-optimized efficiency for daily use** |
|
|
| This makes visual search something you can actually use **every day**, not just when plugged in. |
|
|
| --- |
|
|
| ## Use Cases |
|
|
| - **Personal image libraries:** Rediscover memes, screenshots, and old photos instantly. |
| - **Creative workflows:** Search moodboards and visual references with natural language. |
| - **Edge & embedded systems:** Efficient multimodal search for mobile, XR, IoT, and automotive. |
|
|
| --- |
|
|
| ## Performance Highlights |
|
|
| - Sub-second search even across large image libraries |
| - ~10× lower power consumption vs CPU/GPU search |
| - Stable always-on indexing without thermal or battery issues |
|
|
| ## License |
| This model is released under the **Creative Commons Attribution–NonCommercial 4.0 (CC BY-NC 4.0)** license. |
| Non-commercial use, modification, and redistribution are permitted with attribution. |
| For commercial licensing, please contact **dev@nexa.ai**. |
