- Description
- Specifications
- Download
- FAQ
- Reviews
Overview
InHand Networks Mo 62A AI Single Board Computer
The Mo 62A is an edge-AI single board computer built on the TI AM62A74 vision processor. With a dedicated C7x DSP and deep-learning accelerator delivering 2 TOPS of on-device inference, an on-chip ISP engineered for real-world camera scenes, and a standard SBC form factor compatible with the HAT ecosystem, the Mo 62A turns AI vision ideas into deployable edge devices — at a fraction of the cost and power of traditional AI hardware. Running Debian Linux with an open SDK and TI TIDL support for TFLite/ONNX models, it is built for developers who want to go from prototype to production — with the complete developer environment, tools and examples published on GitHub, the open-source way.
Features
2 TOPS of Edge AI — Inference Where the Camera Is
A dedicated C7x DSP with Matrix Multiply Accelerator delivers 2 TOPS of deep-learning performance on-device. Detection, classification and segmentation models run locally — no cloud round-trips, no bandwidth bills, no privacy exposure. TI TIDL toolchain deploys your existing TFLite and ONNX models with minimal effort.
A Vision Processor, Not Just an SBC
Unlike general-purpose boards, the AM62A integrates an on-chip ISP with VPAC vision acceleration — hardware WDR, lens distortion correction and RGB-IR support. Backlit entrances, high-contrast construction sites and low-light scenes are handled in silicon before your model ever sees the frame, lifting real-world detection accuracy.
Raspberry Pi Form Factor, Industrial AI Core
Standard 85 × 56 mm SBC footprint with a 40-pin HAT-compatible header (GPIO / I²C / SPI / UART / PCM). Mounts in mainstream SBC enclosures and works with the vast ecosystem of HAT expansion boards — PoE HATs, relays, sensors, displays — so prototyping hardware is never a blocker.
Built for Developers — Everything on GitHub
A complete developer environment is published on GitHub, the open-source-community way: ready-to-flash Debian system images, the full SDK for custom builds, AI model deployment examples, pinout and device-tree documentation — clone, flash and run your first inference in minutes. Issues and contributions are handled in the open, so answers stay searchable for everyone. No proprietary runtime in the way: develop in Python or C/C++ with OpenCV, GStreamer and NumPy, manage packages with apt, and access the board over SSH. Your code, your system, your product.
Connectivity for Real Deployments
Gigabit Ethernet for reliable backhaul, four USB 2.0 ports for peripherals and additional cameras, micro HDMI for local display, dual-band Wi-Fi 5 and BLE 4.2 for cable-free installs, and a 4-lane MIPI CSI-2 interface for direct camera input — everything an edge vision terminal needs, nothing it doesn't.
Secure by Design
Secure boot, Arm TrustZone, OP-TEE trusted execution and a hardware AES-256 crypto accelerator protect firmware, models and data end to end — essential when devices are deployed unattended in public spaces and your AI model is your IP.
From Prototype to Production
USB-C 5V/5A power, 25 W maximum consumption and fanless-friendly thermals keep deployment simple and silent. Three memory options (2GB / 4GB / 8GB) let you match cost to workload — validate on a 4GB board today, order production volumes tomorrow, on the same platform and the same software.
Why Mo 62A — Compare Before You Build
General-purpose SBCs leave AI to add-ons; high-end AI kits cost several times more. The Mo 62A puts a dedicated vision AI pipeline in a standard SBC form factor — right-sized for single- and dual-camera edge inference.
| Specification | Raspberry Pi 5 | Jetson Orin Nano |
Typical RK3588 SBC | InHand Mo 68A |
InHand Mo 62A |
|---|---|---|---|---|---|
| Positioning | General-purpose SBC | High-perf AI dev kit | Multimedia SBC | Multi-camera vision AI | Edge AI vision SBC |
| AI acceleration | — (requires AI HAT) | Up to 40 TOPS | 6 TOPS NPU | 8 TOPS | 2 TOPS (C7x DSP + MMA) |
| AI toolchain | Via HAT vendor | TensorRT | RKNN | TIDL · TFLite / ONNX | TIDL · TFLite / ONNX |
| CPU | 4× Cortex-A76 @2.4GHz | 6× Cortex-A78AE | 8× (A76+A55) | 2× Cortex-A72 @2.0GHz | 4× Cortex-A53 @1.4GHz |
| Camera input | 2× CSI (mini) | Via carrier board | MIPI CSI | 2× 4-lane MIPI CSI-2 | 1× 4-lane MIPI CSI-2 |
| Vision ISP | Basic | Yes | Yes | VPAC + DMPAC | ISP + VPAC (WDR · LDC · RGB-IR) |
| Video | 4K60 decode | 4K60 | 8K decode | 4K60 encode/decode | micro HDMI out |
| Expansion | 40-pin GPIO | 40-pin GPIO | 40-pin GPIO | 40-pin HAT-compatible | 40-pin HAT-compatible |
| OS | Raspberry Pi OS | JetPack (Ubuntu) | Linux / Android | Debian 13 | Debian 13 · open SDK |
| Power input | USB-C 5V/5A | 7–20V DC | 12V DC | DC in | USB-C 5V/5A · 25W max |
| Typical budget | $ (board only, no AI) | $$$$ | $$ | $$ | $–$$ |
* Competitor specifications are collected from public sources for reference only. Please refer to each vendor's official documentation for final data. Need 8 TOPS, dual cameras and 4K60 codec? Step up to the Mo 68A.
Applications
From a developer's desk to a thousand deployed sites — one compact platform for vision AI at the edge.
Developer Resources — Open Source on GitHub
We develop in the open. The full Mo 62A developer environment — system images, SDK, AI examples and documentation — is on its way to GitHub, so getting started will be a clone away and every answer stays searchable.
System images · EdgeAI SDK · TIDL model examples (TFLite / ONNX) · GPIO & device-tree docs · Issue tracker — contributions and pull requests will be welcome.
In the Box
Why InHand
InHand Networks has built industrial connectivity and edge computing devices for over two decades, deployed in energy, transportation, manufacturing and smart-city projects worldwide. The Mo series brings that industrial engineering discipline to edge AI: documented hardware, maintained software, and a supply chain you can plan a product on.
Hardware
| Processor | TI AM62A74, 4 × Arm Cortex-A53 @ 1.4 GHz |
| AI Accelerator | C7x DSP + Deep Learning Accelerator (MMA), 2 TOPS |
| ISP / Vision | On-chip ISP + VPAC (RGB-IR, WDR, LDC) |
| RAM | LPDDR4 4GB (default) / 2GB / 8GB |
| Storage | Micro SD card |
| Ethernet | 1 × Gigabit Ethernet |
| USB | 4 × USB 2.0 Type-A |
| Display | 1 × micro HDMI |
| Camera | 1 × 4-lane MIPI CSI-2 |
| Audio | 3.5 mm jack + PCM (via 40-pin connector) |
| Expansion | 40-pin header: GPIO / I²C / SPI / UART / PWM / PCM, 3.3 V logic, HAT-compatible |
| Debug | 1 × TTL UART console |
| Fan | 1 × 4-pin fan header (active cooling optional) |
| Button | 1 × Reset |
| LED | PWR, USER |
| RTC | Supported, with battery backup |
Wireless
| Wi-Fi | Wi-Fi 5 (802.11ac), dual-band |
| Bluetooth | BLE 4.2 |
| Antenna | Onboard snap-on antenna (Wi-Fi / BLE) |
Power & Mechanical
| Power Input | USB Type-C, 5V / 5A DC |
| Power Consumption | 25 W (max) |
| Dimensions | 85 × 56 mm |
| Weight | 47 g |
| Operating Temperature | 0 ~ 50 °C |
| Storage Temperature | -20 ~ 70 °C |
Software
| Operating System | Debian 13.2 (Trixie), Linux kernel 6.12 |
| AI Runtime | TI TIDL — deploys TFLite / ONNX models |
| Vision SDK | TI EdgeAI SDK; V4L2 camera framework; DRM/KMS display |
| Development | Python, C/C++; OpenCV, GStreamer, NumPy; apt package manager |
| Security | Secure boot, Arm TrustZone, OP-TEE, hardware AES-256 |
| Open SDK | Custom system builds supported |
| Management | SSH remote access; SD card image flashing; UART console |
| Networking | TCP/UDP, ICMP, DNS, DHCP; static routing |
Models
| Model | RAM |
|---|---|
| Mo-62A-2G | 2GB LPDDR4 |
| Mo-62A-4G | 4GB LPDDR4 |
| Mo-62A-8G | 8GB LPDDR4 |
The complete developer environment is coming to GitHub: latest system images, SDK source, TIDL model examples and docs. Follow our page for the launch announcement.
What does the Mo 62A offer?
The Mo 62A is a 2-TOPS AI single-board computer for on-device vision inference at the edge. Key specs:
| Compute | 4× Cortex-A53 @ 1.4 GHz + C7x DSP + 2-TOPS deep-learning accelerator |
| Vision pipeline | On-chip ISP + VPAC (RGB-IR, WDR, LDC); 1× 4-lane MIPI CSI-2 |
| RAM | LPDDR4 — 2 GB / 4 GB (default) / 8 GB SKUs (Mo-62A-2G / -4G / -8G) |
| OS | Debian Linux |
| Connectivity | 1× Gigabit Ethernet, Wi-Fi 5 dual-band, BLE 4.2 |
| Other I/O | 4× USB 2.0 Type-A, micro HDMI, 3.5 mm audio, 40-pin HAT-compatible header, micro SD |
| Security | Secure Boot, Arm TrustZone, OP-TEE, hardware AES-256 |
| Power / size | USB Type-C 5 V (25 W max), 85 × 56 mm, 47 g |
The 85 × 56 mm board and 40-pin layout are mechanically compatible with standard SBC enclosures and HAT accessories. Typical applications: AI vision boxes, intelligent cameras, defect inspection and other edge-AI inference terminals. The Mo 62A is a developer-oriented SBC managed via SSH and standard Debian tooling — it does not run InHand IEOS or DeviceLive.
Which AI frameworks and model formats are supported?
Models run through the TI Deep Learning (TIDL) runtime, which accepts TFLite (.tflite) and ONNX (.onnx) models. The on-chip ISP/VPAC handles RAW→RGB conversion, WDR and lens correction so the accelerator gets a clean input; pre- and post-processing in Python or C/C++ commonly use OpenCV and GStreamer. Ready-to-run model examples are published on our GitHub.
What workloads is 2 TOPS enough for?
At 2 TOPS the Mo 62A is sized for image classification, lightweight object detection (e.g. YOLO-Nano variants), OCR, presence/anomaly sensing, and gesture or pose recognition. For multi-stream or high-resolution detection at higher frame rates, step up to the 8-TOPS InHand Mo 68A.
How do I use the 40-pin GPIO header?
The 40-pin HAT-compatible header exposes GPIO, I²C, SPI, UART, PWM and PCM at 3.3 V logic. Standard SBC HAT accessories mount mechanically — verify electrical compatibility (3.3 V) before connecting.
Pin notes: all user pins default to GPIO, with alternate functions enabled via device-tree overlays; pins 27/28 are reserved for the camera I²C bus and cannot be used as general GPIO. The board ships with libgpiod v2.x — the -c flag is required to specify a chip, and three gpiochips (gpiochip0–gpiochip2) cover the MCU and main domains. Quick examples:
gpiodetect # list chips gpioget -c gpiochip2 23 # read pin 11 gpioset -c gpiochip1 39=1 # drive pin 7 high gpiomon -c gpiochip2 23 # watch edges i2cdetect -y 2 # scan camera I²C bus
How do I connect cameras, displays and high-bandwidth peripherals?
Use the 4-lane MIPI CSI-2 connector for direct camera input — the IMX219 is supported out of the box via the bundled imx219-preview.sh script. The micro HDMI drives local displays, and the four USB 2.0 Type-A ports handle storage, adapters and additional cameras.
Does it work with Raspberry Pi HATs and enclosures?
Mechanically, yes — the 85 × 56 mm footprint and 40-pin layout match the standard SBC form factor, so mainstream enclosures and HATs mount directly. Electrically, the header is 3.3 V logic: check your HAT's voltage requirements before connecting.
No reviews yet — be the first to share your Mo 62A project experience.
