NeoEyes NE302 Puts 600 GOPS of Vision AI on USB-C Power

A $99 Edge AI Camera That Runs on Nothing but USB-C

CamThink has launched the NeoEyes NE302, a compact edge AI vision node built around an STM32N6 microcontroller. It sells for $99 at launch against a $109 list price, measures 42 x 42 x 20 mm in its enclosure, and draws all of its power from a single 5V USB-C connection. There is no fan, no battery, and no separate AI accelerator board — the inference happens on the microcontroller itself.

• STM32N6 with an Arm Cortex-M55 at up to 800 MHz and an ST Neural-ART NPU rated at 600 GOPS
• 4 MP OS04C10 sensor with a choice of 88-degree wide or 137-degree ultra-wide lens
• 32 MB PSRAM, 4.2 MB SRAM, 64 MB SPI flash, plus a microSD slot
• Wi-Fi 6 and Bluetooth 5.3 LE; $99 launch price, $109 MSRP; 42 x 42 x 20 mm

Why 600 GOPS on a Microcontroller Is the Interesting Number

ST quotes the Neural-ART accelerator at up to 600 GOPS with an efficiency figure of 3 TOPS per watt, and that efficiency is what makes the fanless enclosure possible. Plenty of single board computers will beat 600 GOPS on paper; very few will do it inside a 20 mm-thick sealed box with no moving parts and a USB-C cable as the only service requirement.

The practical envelope is object detection and pose estimation running locally, with a hardware video path that takes MIPI CSI-2 into the ISP and out as H.264 or MJPEG — up to 1920x1080 at 30 fps according to CamThink's product page. That combination is the point. You get a detection result and an encoded stream from the same 5V supply, without a Linux box in the loop.

For context on how quickly this class of silicon has moved, compare it to the 8 TOPS edge AI modules we looked at in the Boardcon CM311Y3 system-on-module. Those are application processors with DRAM and an operating system to maintain. The NE302 is a microcontroller that boots in a blink and does one job.

What Did CamThink Remove From the NE301?

This is a leaner sibling rather than a straight upgrade, and it is worth being clear about the trade. Relative to last year's NE301, the NE302 drops the battery, the 4G LTE modem, audio connectors, the PIR motion sensor connector, the 16-pin GPIO header, and Power over Ethernet. Memory comes down to 32 MB PSRAM and 64 MB SPI flash.

What you gain is Wi-Fi 6, a smaller footprint, a lower price, and a much simpler power story. If your deployment has mains power and a wireless network — a kiosk, a retail display, a production line station, a workshop bench — the removed parts were cost you were not using. If you need a camera on a pole in a field, the NE301 with its battery and LTE is still the right board.

Integration and the Open-Source Angle

The NE302 speaks a sensible set of protocols out of the box: a web UI, an HTTP REST API, WebSocket, MQTT, RTSP, and RTMP. MQTT in particular means it drops into an existing home or factory automation setup without glue code, in much the same way the HomeMaster MiniPLC brought ESPHome to a DIN rail.

CamThink publishes firmware, documentation, and schematics on GitHub and its own wiki, and the product page invites developers to bring their own models or pick from the company's model zoo. The board-level variant is offered for integration into other products, though OEM pricing has not been disclosed. Operating range is a fairly ordinary -20°C to 50°C, which reads as indoor and sheltered-outdoor rather than industrial.

Should You Buy One?

At $99, the NE302 is priced where experiments stop needing justification. If you have wanted a camera that classifies what it sees before anything leaves the device — for privacy reasons, bandwidth reasons, or simply because the network is unreliable — this is one of the cheapest serious entry points available. Our mini computer coverage has watched local inference migrate steadily down from mini PCs to modules to microcontrollers, and a fanless 42 mm cube doing pose estimation on bus power is a good marker of where that line now sits.

One caveat on sourcing: the launch specifications here come from CamThink's own product page and from CNX Software's September 12 report, which is currently the only independent outlet to have covered the NE302 in detail. Benchmarks from third parties are not yet available, so treat the 600 GOPS figure as the silicon vendor's rating rather than a measured result.

Sources: CNX Software — September 12, 2026; CamThink product page — September 2026.