loading . . . MaixCAM2 modular 4K AI camera is based on Axera AX630 SoC with 3.2 TOPS NPU (Crowdfunding) Designed for makers and researchers, Sipeed’s **MaixCAM2** is a modular AI camera built around an Axera AX630 AI SoC with 3.2 TOPS (INT8) to 12.8 TOPS (INT4) of AI performance, yet delivering 10 to 20x higher performance on OpenMV-N6 and Raspberry Pi 5, and about the same performance (140 FPS) as a 33 TOPS Jetson Orin Nano 4GB Super in YOLO11n.
The camera comes with 1GB or 4GB RAM, a 32GB eMMC flash, and a microSD card slot for storage, a 2.4-inch touchscreen display, and an 8MP / 4Kp30 camera sensor with LED flash, as well as two microphones and a speaker, making it suitable for local LLM and VLM applications. It also features two PMOD connectors for expansion, such as a thermal imager and a Time of Flight (ToF) ranging sensor.
MaxiCAM2 specifications:
* SoC – Axera Tech AX630
* CPU
* Dual-core Arm Cortex-A53 @ 1.2 GHz running Ubuntu Linux
* 32-bit RISC-V E907 core running RTT
* NPU – 12.8 TOPS @ INT4, 3.2 TOPS @ INT8; supports convolution and Transformer models such as YOLO/LLM/VLM, YOLO11n 640×640 reaches up to 113FPS
* ISP – 4K @ 30fps
* VPU – H.264 / H.265 / MPJEG; encoding: 4K @ 30 fps; decoding:1080p @ 60fps
* System memory – 1GB or 4GB LPDDR4
* Storage
* 32GB eMMC 5.1 flash
* MicroSD card slot
* Display
* 2.4-inch HD (640×480) IPS capacitive touchscreen
* 4-lane MIPI DSI output, 31-pin interface, 6-pin capacitive touchscreen, up to 1080p60 output
* Camera
* Up to 8MP 4K @ 30fps
* 1/1.8″ image sensor (up to 80% larger than the Raspberry Pi Camera, and 10–20% larger than sensors found in popular commercial cameras like Insta360 and GoPro)
* M12 lens, convertible to a CS-mount for industrial cameras.
* Extra screw fixing holes allow for the installation of a 1.25-inch adaptor for astronomical telescopes or a 23.2mm adaptor for microscopes.
* 4-lane MIPI CSI input, 22-pin interface
* Audio
* Onboard PA amplifier + 1W speaker
* 2x analog silicon mics, direct audio capture
* Networking
* WiFi 6 and BLE 5.4 module
* 6-pin FPC Ethernet interface (with external FPC-to-RJ45 module)
* USB – 1x USB 2.0 Type-C port (Device and Host modes), supports USB cameras
* Sensor – Onboard 6-axis IMU (3-axis accelerometer + 3-axis gyroscope)
* Expansion
* 2x 2.54mm pitch PMOD interfaces for expansion (20x IOs + Vsys/3.3V/GN)
* 6-pin 1.25mm pitch expansion connector
* Those support I2C, SPI, UART, ADC, PWM, WDT, and other common peripherals
* Misc
* Power switch
* Func (function) button
* Power, User, and Illumination LEDs
* Onboard BM8563EMA RTC chip + rechargeable coin battery
* Power Supply
* 5V via USB-C port
* Li-ion battery with charge/discharge management
* Power Consumption – 2.5 Watts in Qwen3-VL-2B (excluding sensor power consumption)
* Dimensions – 65 x 49 x 20mm with protective case; includes 1/4-inch standard tripod mount
The MaixCAM2 has been designed to be easy for beginners, with over 40+ AI apps preinstalled, including Qwen3-VL-2B VLM, Yolov11n AI vision model, and training-free AI Models such as YoloWorld and Mixformerv2. Additional apps can easily be installed from the MaixHub App store.
Advanced users can leverage the MaixPy Python environment and the MaixVision IDE to create their own models. More details about the MaixCAM2 hardware, performance, and MaixPy development can be found on the wiki and GitHub.
As mentioned in the introduction, the camera is modular thanks to its two PMOD connectors for a thermal camera (32×24, 160×120, or 256×192, and up to 640×480 with AI super resolution) and/or a ToF sensor up to 100×100 resolution.
Sipeed compared the MaixCAM2’s 4K 1/1.8-inch sensor against competitors like the Raspberry Pi Camera Module 3, action cameras, smartphone cameras, Link2, and ASP-C cameras under the same low-light conditions in the photos below.
MaixCAM2 comparison against Raspberry Pi Camera Module 3, action cameras, mobile phone cameras, etc…
They also shared YOLO11n model benchmarks against other platforms like the Raspberry Pi 5, Rockchip RK3588, and NVIDIA Jetson Orin Nano 4GB (at 33 TOPS, not 67 TOPS).
Qwen3-VL-2B benchmark and efficient data were also provided. The MaxCAM2 can’t quite match the performance of an AMD 8845HS, but it’s much more efficient. It’s also three times faster than a Raspberry Pi 5.
It’s at least the fourth device with an AX630/AX630C we’ve come across, after M5Stack Module LLM and LLM630 Compute Kit, as well as Sipeed’s own NanoKVM Pro 4K IP KVM. It’s just the first time we see a ready-to-use AI camera based on it.
The video embedded at the end of this post has a few demos showcasing the capabilities of the MaixCAM2 4K AI camera. Sipeed has just launched it on Kickstarter with a 50,000 HKD (about $6,400 US) funding target. Rewards start at $69 for the MaixCAM2 with 1GB RAM and a 2K camera sensor, $99 for the 4GB/2K variant, and $119 for the 4GB/4K model highlighted in this post. Shipping adds 40 HKD to Hong Kong, and $15 to $25 US to the rest of the world. Deliveries are scheduled to start at the end of next month, right after the crowdfunding campaign is over.
Jean-Luc Aufranc (CNXSoft)
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.
Share this:
**Support CNX Software! Donate viacryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress**. We also use affiliate links in articles to earn commissions if you make a purchase after clicking on those links. https://www.cnx-software.com/2026/01/27/maixcam2-modular-4k-ai-camera-is-based-on-axera-ax630-soc-with-3-2-tops-npu/
