CamThink NeoEyes NE302: STM32N6 Neural-ART Edge AI Vision Board

What is the NeoEyes NE302 & why did it catch our eye?

The NeoEyes NE302 is a thumb-sized edge AI vision camera board built by CamThink around STMicroelectronics' new STM32N6 neural microcontroller. It runs full neural object detection and classification inside an ultra-low-power MCU envelope with zero cloud dependencies.

Most computer vision projects hit a frustrating wall: you either deploy a power-hungry Single Board Computer like a Raspberry Pi 5 running a full Linux OS with thermal throttling, or you use a low-power ESP32 that chokes when running anything heavier than simple blob detection. The NeoEyes NE302 changes that equation.

When our team at ZeroShot Studio analyzed the STM32N6 architecture, what stood out was the dedicated Neural-ART accelerator. ST paired an 800 MHz Arm Cortex-M55 core with a 1 GHz proprietary NPU capable of 600 GOPS at an astonishing 3 TOPS/W efficiency. This means you get real-time vision inference in a fanless, battery-friendly form factor that draws less than 2 Watts under full load.

Hardware specs at a glance

Here is how the silicon and I/O stack up under the hood:

The DIY factor: Can you actually hack and mod it?

For DIY builders, the biggest trap with modern camera modules is closed firmware. Many vendors sell hardware that only speaks to a proprietary mobile app. The NeoEyes NE302 avoids that trap:

1. Open toolchain support: Firmware development is grounded in the official STM32CubeN6 BSP and CMSIS-NN neural libraries. You can compile with GCC and flash directly without paid vendor toolkits.
2. Local network protocols: Out of the box, the firmware exposes an RTSP video stream, a lightweight HTTP REST configuration API, a WebSocket event bus, and MQTT telemetry publishing.
3. Dedicated STLINK debugging: Unlike throwaway smart cameras, CamThink routed SWD debug pads directly to standard header positions. If you wipe the flash or hang the Cortex-M55 core during bare-metal driver development, an ST-Link V3 recovers the device immediately.

flowchart TD
    subgraph VisionEngine ["Edge Sensor & Neural Engine"]
        direction LR
        Sensor["OmniVision 4MP Sensor<br />OS04C10 · MIPI CSI-2"] --> NPU["STM32N6 Neural-ART NPU<br />1 GHz / 600 GOPS"]
        NPU -->|Bounding Boxes & Tensors| MCU["Cortex-M55 MCU<br />800 MHz Host"]
    end

    Broker["Local MQTT Broker / ZeroMini<br />Wi-Fi 6 RTSP & MQTT"]
    Agent["OpenClaw Workshop Agent<br />JSON Event Stream Trigger"]

    MCU -->|Telemetry & Video Stream| Broker
    Broker -->|JSON Event Stream| Agent

How does it hook into coding agents and vibe coding?

Because the NE302 speaks standard MQTT and HTTP, you do not need custom native C drivers on your server to interface with it. Coding agents like Claude Code, Cursor, and OpenClaw can generate Python automation workers in seconds.

Here is a minimal telemetry monitor that subscribes to real-time object classification events emitted by the board:

# Minimal telemetry listener for NeoEyes NE302import jsonimport paho.mqtt.client as mqttNE302_IP = "192.168.1.142"TOPIC_DETECTIONS = "neoeyes/telemetry/detections"def on_connect(client, userdata, flags, rc):    print(f"[connected] Subscribed to {TOPIC_DETECTIONS} (code {rc})")    client.subscribe(TOPIC_DETECTIONS)def on_message(client, userdata, msg):    payload = json.loads(msg.payload.decode("utf-8"))    detections = payload.get("objects", [])    for item in detections:        label = item.get("label")        confidence = item.get("confidence", 0.0)        box = item.get("bbox", [])        print(f"[vision-alert] Detected {label} ({confidence:.1%}) at {box}")if __name__ == "__main__":    client = mqtt.Client(client_id="openclaw-vision-node")    client.on_connect = on_connect    client.on_message = on_message    client.connect(NE302_IP, 1883, 60)    client.loop_forever()

What is the catch before you order one?

No hardware release is without trade-offs:

• Supply chain and early silicon: The STM32N6 is cutting-edge silicon launched in late 2026. Lead times from distributors like Mouser and DigiKey may fluctuate during initial production runs.
• Model conversion toolchain: Converting PyTorch or ONNX models to the proprietary Neural-ART NPU format requires ST's STM32Cube.AI model optimizer. While the runtime is royalty-free, the conversion pipeline takes some trial and error with quantization layers.
• Enclosure design: The bare board comes as a PCB sandwich with exposed CSI ribbon cables. You will want to 3D print an enclosure to protect the optics and thermal pads before workshop deployment.

What we are building with it next

We are staging the NE302 on our workshop test bench. Planned explorations include:

1. Local tripwire sentinel: Running a custom YOLOv8-nano model trained on tool shapes to automatically verify that workshop safety gear is in place.
2. OpenClaw hardware trigger: Pairing the NE302's MQTT event stream to autonomous agent tasks, triggering context summaries when someone approaches the desk.
3. Benchmarking with ESP-Mosaico: Comparing inference latency and energy efficiency against our modular ESP-Mosaico board.

Check out our full collection of maker guides in the ZeroLabs Resources directory.

FAQ

Where can I buy the CamThink NeoEyes NE302? The board is available through maker electronics distributors and directly from CamThink and STMicroelectronics distribution channels.

Can I run custom neural networks on the STM32N6? Yes. ST's STM32Cube.AI toolchain ingests ONNX, TensorFlow Lite, and Keras models, quantizing them into optimized C kernels for the Neural-ART NPU.

Does the board require a cloud subscription? No. All neural inference, video encoding, and network services run 100% locally on the device with zero cloud connectivity required.