The OpenMV AE3, a 1-inch-square AI camera board, funded on Kickstarter alongside its sibling board the N6, with backers pre-ordering the AE3 for $80 and shipping planned before year’s end. What makes it notable isn’t just the size — it’s what fits inside that size: dual Arm Cortex-M55 MCU cores (400MHz primary, 160MHz secondary) paired with dual Arm Ethos-U55 microNPU cores (256 MACs and 128 MACs respectively), enough dedicated neural-processing silicon to run real object-detection models directly on the board.
Watch: OpenMV — Machine Vision Camera Module Review (YouTube)
Real object detection, not a toy demo
The AE3 supports models from Google’s MediaPipe, along with YOLOv2, YOLOv5, and YOLOv8 — meaning it can run genuine, production-grade object-detection architectures rather than a stripped-down edge-AI demo. In practice, that means running YOLO-based detection at roughly 30 frames per second while drawing only 0.25 watts of power, a combination of real-time performance and power efficiency that’s historically required much larger, more power-hungry hardware.
The power numbers are the real story
Full-power draw sits under 60mA at 5V — 0.25W — and deep sleep drops to under 2.5mW, low enough that a battery-powered AE3 running an intermittent detection workload can realistically operate for years on a single charge rather than days. Connectivity includes Wi-Fi and Bluetooth, a USB-C port, and a board-to-board connector, alongside a built-in microphone, accelerometer, time-of-flight sensor, RGB LED, and a Qwiic connector for direct expansion into SparkFun’s sensor ecosystem — a board built to be a complete, deployable sensing node on its own, not just a camera module needing a host computer nearby.
Why MicroPython matters for this specific board
The AE3 is programmed in MicroPython rather than requiring a C/C++ embedded toolchain, which matters enormously for who can actually build with it: a maker prototyping a wildlife camera trap, an interactive installation that reacts to specific detected objects, or a generative-art piece driven by real-time object counts can iterate in a scripting language rather than fighting an embedded build system, while still getting genuine on-device neural inference rather than streaming frames to a cloud API.
Related Reading
- OpenMV N6 & AE3 — Low Power, Python Programmable AI Cameras — Kickstarter
- Introducing the OpenMV AE3 and N6 Cameras — Hackster.io
- OpenMV AE3 edge AI cam packs Alif Ensemble E3, dual NPUs, 1″ global-shutter vision, YOLO at the edge — ARMdevices.net
- MicroPython-programmable OpenMV N6 and AE3 AI camera boards run on battery for years — CNX Software