Available

MT200 AI Box

Industrial-grade edge AI box with up to 200 TOPS of local compute, built on Qualcomm QCS9075 and the MeiG SNM983 module. Runs large models on-site — no cloud required.

Overview

An AI workstation for the factory floor

MT200 is a modular edge AI computer. The flagship configuration pairs the MeiG SNM983 module — based on Qualcomm QCS9075 (IQ-9 series, 5 nm) — with an 8-core Kryo Gen 6 CPU, Adreno 663 GPU and dual Hexagon Tensor Processor NPUs.

It supports 4K multi-display output, concurrent camera input and 8K video decoding, with industrial interfaces including CAN-FD, PCIe, GPIO and dual Ethernet. Typical deployments include intelligent robotics, industrial vision, in-vehicle systems and edge gateways.

  • 100 TOPS dense / 200 TOPS sparse INT8
  • 12 / 24 / 36 GB LPDDR5, 128 GB UFS 3.1
  • Ubuntu 24.04 LTS or Linux Yocto
  • −40 °C to +85 °C operating range
MT200 AI Box front panel with dual Ethernet, four USB-A ports, HDMI input and output, and DC power
Highlights

Built for AI agents, not just inference

COMPUTE

200 TOPS, dual NPU

Two Hexagon Tensor Processor DSPs let multiple models run concurrently — vision, speech and LLM in the same pipeline.

MODELS

7B parameters on-device

Qwen, LLaMA and DeepSeek-derived models run locally; no API key and no per-token billing.

MULTIMODAL

See, hear, speak

Vision (YOLO, Qwen2-VL), speech recognition (Whisper) and TTS (MeloTTS) on one box.

AGENTS

Agent-ready runtime

OpenClaw validated on MT200, with chat-app integration for Feishu and DingTalk and GPIO control.

PRIVACY

Data never leaves

Inference, storage and logs stay on the device. Suitable for regulated and air-gapped environments.

INDUSTRIAL

Wide temperature, real I/O

−40 to +85 °C, 12/24 V input, CAN-FD, PCIe Gen 4, dual Ethernet and a Raspberry Pi header.

Hardware

125 × 100 × 53 mm

Aluminium enclosure with active cooling, front-facing I/O and a serviceable heatsink module.

Specifications

Two configurations

MT200_8 covers 24–60 TOPS platforms; MT200_9 is the flagship QCS9075 configuration.

Model comparisonMT200_8MT200_9
Compute platformSNM979 (CQ7790S) 24 TOPS
SNM970 (QCS8550) 48 TOPS
SNM982 (QCS8625) 60 TOPS
SNM983 (QCS9075, IQ-9) 100 TOPS dense / 200 TOPS sparse
Operating systemLinux / Ubuntu*Linux Yocto / Ubuntu 24.04 LTS
Ethernet2 × RJ452 × RJ45 (2.5 GbE + 1 GbE)
Display output1 × HDMI out1 × HDMI out, 1 × DP
Display input1 × HDMI in1 × HDMI in
Storage slotM.2 Key B&M, 2-lane PCIe Gen 4M.2 Key M, 4-lane PCIe Gen 4 (128 GB UFS onboard)
Wireless / cellularWi-Fi module (M.2 Key E); data module (M.2 Key B)Wi-Fi module (M.2 Key E) or data module (M.2 Key B) — one at a time
USB1 × USB-C with DP 1.4, 4 × USB-A (QPS615)1 × USB-C 3.0, 4 × USB-A 3.0 (hub)
microSD1 × microSD—
Audio3.5 mm jack3.5 mm jack
GPIO headerRaspberry Pi 40-pinRaspberry Pi 40-pin
CAN—2 × 4-pin CAN FD
Camera2 × D-PHY2 × D-PHY

* Ubuntu availability depends on the selected compute module. Specifications are subject to change; confirm the final configuration with us before design-in.

MT200_9 details

CategorySpecification
AI performance100 TOPS dense INT8 / 200 TOPS sparse INT8
AI frameworksONNX (CPU/GPU), TensorFlow Lite (CPU/GPU), QNN, SNPE
CPU8-core Kryo Gen 6, up to 2.36 GHz
GPUAdreno 663 @ 800 MHz
NPUDual Hexagon Tensor Processor DSP @ 1.42 GHz
Audio DSPHexagon Audio DSP @ 1.344 GHz
Memory12 GB / 24 GB / 36 GB LPDDR5, 3200 MHz
Storage128 GB UFS 3.1 Gen4, 2-lane
Video decode4K @ 275 fps — AV1, H.264, H.265, VP9, MPEG-2
Video encode4K @ 170 fps — H.264, H.265, HEIF/HEIC
Camera interface2 × D-PHY (internal adapter board)
USB4 × USB 3.0 Type-A (host), 1 × USB 3.0 Type-C
PCIe expansionM.2 Key M 4-lane PCIe Gen 4; M.2 Key E or Key B 2-lane PCIe Gen 4
DebugUART through-hole pad
PowerExternal 12 V / 24 V adapter
RTCCR2032 (225 mAh)
ButtonsPower, volume +, volume −, EDL (pinhole)
Dimensions125 × 100 × 53 mm
Operating temperature−40 °C to +85 °C
Storage temperature−40 °C to +105 °C
CertificationsRoHS, EMC, UL (some items under development)

Source: MeiG MT200 specification sheet (2026-07).

Software

Models validated on MT200

Bring your own model or start from this list. Most run through ONNX or the Qualcomm QNN backend.

Large language models

Qwen2 0.5BQwen2.5 3BQwen2.5 7BQwen3 0.6BQwen3 1.7BLLaMA 2 7BDeepSeek-Qwen 1.5BDeepSeek-Qwen 7BDeepSeek-LLM 7B

Vision-language & generation

Qwen2-VL 2BQwen2-VL 7BStable Diffusion 1.5Stable Diffusion 2.1Depth-AnythingMiDaS v2

Speech

WhisperHuggingFace WavLMMeloTTSYAMNet

Vision & detection

YOLOv3–v11YOLO-XRTMDetDETR ResNetHRNetPoseMediaPipe Pose / Hand / FacePPE DetectionPerson / Foot DetectionSwin TransformerViT
Deployments

What people build with it

VISION + GPIO

Natural-language hardware control

Ask in chat, and the local VLM reads the schematic or camera feed, then drives GPIO to actuate equipment.

INDUSTRIAL VISION

Line inspection and safety

YOLO-class detectors run locally on camera streams for defect detection, PPE checks and access monitoring.

AUTOMOTIVE

In-vehicle compute

Multi-display output, CAN-FD and wide-temperature operation suit telematics and cabin intelligence.

PRIVATE AI

Air-gapped assistants

Document parsing, summarisation and agent workflows run without any external network connection.

MT200 is designed and manufactured by MeiG Smart Technology Co., Ltd. (SZSE: 002881). TailAether provides product selection, configuration guidance, integration support and delivery. Specifications and availability follow MeiG Smart's official documentation.

Need a configuration for your workload?

Tell us the model, camera count, I/O and environment. We will propose a variant and quote.

Request a quote