XINK SOM
The drop-in robot brain
- Compute
- eCV SoC — quad A55 @ 1.8 GHz + M4
- NPU
- 4.1 TOPS (@900 MHz)
- Memory
- 4 GB LPDDR4 · 32 GB eMMC
- Video
- H.264 / JPEG codec
- Size
- 40 × 55 mm
← All products
reason
XINK COREShips now
The reasoning core of Physical AI.
XINK CORE enables edge devices and robots to understand context, reason locally, and generate intelligent actions without relying fully on cloud AI — from the eCV AI SoC shipping today to a dedicated LLM processor on the roadmap.
Compute modules & dev kits
Server-class AI capability in space- and power-efficient modules — drop in the core board, or start from a full development kit.
XINK V2
Full-I/O development kit
- Camera in
- 4× MIPI Rx
- USB
- 4× USB 3.0 + 4× USB 2.0
- Display
- HDMI 1080p
- Network
- Gigabit Ethernet + M.2 Wi-Fi
- Power
- 12 V DC
- Size
- 177 × 134 mm
XINK NANO
Compact sensor-fusion platform
- Camera in
- 3× MIPI Rx
- USB
- 2× USB 3.1 A · 2× USB-C
- Display
- Micro-HDMI output
- Network
- Gigabit Ethernet + M.2 Wi-Fi
- Power
- 12 V DC / PoE / USB-C PD
- Size
- 105 × 80 × 32 mm
Reason at the edge
Edge LLM / VLM models run where the data is captured — private, low-latency inference that keeps working when the network doesn't. In live demos, a quantized Gemma 3 model on a XINK board understands natural-language commands, plans multi-step tasks, and drives a robot arm.
The AI controller inside
At the center is the eCV AI SoC: a quad-core Arm Cortex-A55 for applications, an always-on Cortex-M4 for real-time control, and an NPU delivering up to 4.6 TOPS — with MIPI, USB 3.0, and Ethernet interfaces that fuse RGB, ToF, IMU, LiDAR, and thermal sensors on one chip.
Next: a dedicated LLM processor
eYs3D is taping out a next-generation 28 nm LLM IC targeting efficient low-context inference — prefill and decode tuned for the edge. It completes the Sense-and-React architecture: perception and cognition operating as a single, unified edge reflex.
Deployed through the Model Zoo
CORE models are versioned, converted, and shipped to devices through the platform's Model Zoo — from base model to fleet rollout in one pipeline.
XINK NANO POC
| OS | ARM64 Ubuntu Mate 20.04 |
|---|---|
| RGB sensor | STMicro VD66GY · 1124 × 1364 |
| RGB module | eSP776 ISP · up to 1120 × 1120 YUY2 · 100° / 57° FoV |
| RGB lens | NewMax DS-69017 (eYs3D IP) |
| ToF sensor | STMicro VD55H1 · raw 672 × 2420 @ 30 fps |
| ToF depth | 84 × 100 point cloud @ 15 fps · H63.8 × V55.1 · <3 W |
| Storage | microSD · eMMC |
| I/O | 2× 1c4d MIPI in · USB 3.0 Type-C · USB 2.0 Type-A · GbE · M.2 Wi-Fi · HDMI |
| GPIO | 12× GPIO — 2× I²C · 1× SPI · 2× PWM |
| Audio | Built-in mic · speaker header · 3.5 mm jack |
| Power | 12 V DC jack or 5 V USB-C · out 12 / 5 / 3.3 V |
| Size | Box 75 × 80 × 30 mm · PCBA 75 × 60 mm |
Works with the AIoT Platform
Every CORE model is managed fleet-wide: versioned in the Model Zoo, installed over the air, monitored from the dashboard.
XINK CONNECT