# **Overview**

**AutoNeural** is a next-generation, **NPU-native multimodal vision–language model** co-designed from the ground up for real-time, on-device inference. Instead of adapting GPU-first architectures, AutoNeural redesigns both **vision encoding** and **language modeling** for the constraints and capabilities of NPUs—achieving **14× faster latency**, **7× lower quantization error**, and **real-time automotive performance** even under aggressive low-precision settings.

AutoNeural integrates:

* A **MobileNetV5-based vision encoder** with depthwise separable convolutions.
* A **Liquid AI hybrid Transformer-SSM language backbone** that dramatically reduces KV-cache overhead.
* A **normalization-free MLP connector** tailored for quantization stability.
* Mixed-precision **W8A16 (vision)** and **W4A16 (language)** inference validated on real Qualcomm NPUs.

AutoNeural powers real-time cockpit intelligence including **in-cabin safety**, **out-of-cabin awareness**, **HMI understanding**, and **visual + conversational function calls**, as demonstrated in the on-device results (Page 6 figure) .

---

# **Key Features**

### 🔍 **MobileNetV5 Vision Encoder (300M)**

Optimized for edge hardware, with:

* **Depthwise separable convolutions** for low compute and bounded activations.
* **Local attention bottlenecks** only in late stages for efficient long-range reasoning.
* **Multi-Scale Fusion Adapter (MSFA)** producing a compact **16×16×2048** feature map.
* Stable **INT8/16** behavior with minimal post-quantization degradation.

Yields **5.8× – 14× speedups** over ViT baselines across 256–768 px inputs.

---

### 🧠 **Hybrid Transformer-SSM Language Backbone (1.2B)**

Designed for NPU memory hierarchies:

* **5:1 ratio of SSM layers to Transformer attention layers**
* **Linear-time gated convolution layers** for most steps
* **Tiny rolling state** instead of KV-cache → up to **60% lower memory bandwidth**
* **W4A16 stable quantization** across layers

---

### 🔗 **Normalization-Free Vision–Language Connector**

A compact 2-layer MLP using **SiLU**, deliberately **removing RMSNorm** to avoid unstable activation ranges during static quantization.

Ensures reliable deployment on W8A16/W4A16 pipelines.

---

### 🚗 **Automotive-Grade Multimodal Intelligence**

Trained on **10M Infinity-MM samples** plus **200k automotive cockpit samples**, covering:

* AI Sentinel (vehicle security)
* AI Greeter (identity recognition)
* Car Finder (parking localization)
* Passenger safety monitoring

Ensures robust performance across lighting, demographics, weather, and motion scenarios.

---

### ⚡ **Real NPU Benchmarks**

Validated on **Qualcomm SA8295P NPU**:

| Metric                    | Baseline (InternVL 2B) | **AutoNeural-VL** |
| ------------------------- | ---------------------- | ----------------- |
| **TTFT**                  | ~1.4 s                 | **~100 ms**       |
| **Max Vision Resolution** | 448×448                | **768×768**       |
| **RMS Quant Error**       | 3.98%                  | **0.56%**         |
| **Decode Throughput**     | 15 tok/s               | **44 tok/s**      |
| **Context Length**        | 1024                   | **4096**          |

---

# **How to Use**

> ⚠️ **Hardware requirement:** AutoNeural is optimized for **Qualcomm NPUs**.

### 1) Install Nexa-SDK

Download the SDK，follow the installation steps provided on the model page.

---

### 2) Configure authentication

Create an access token in the Model Hub, then run:

```bash
nexa config set license '<access_token>'
```

---

### 3) Run the model

```bash
nexa infer NexaAI/AutoNeural
```

### Image input

Drag and drop one or more image files into the terminal window.
Multiple images can be processed with a single query.

---

# **License**

The AutoNeural model is released under the **Creative Commons Attribution–NonCommercial 4.0 (CC BY-NC 4.0)** license.

You may:

* Use the model for **non-commercial** purposes
* Modify and redistribute it with attribution

For **commercial licensing**, please contact:
**[dev@nexa.ai](mailto:dev@nexa.ai)**