Product spec sheet V1.0

AI BOX

Edge AI computing terminal

Armv9 12-core CPU and 45 TOPS NPU, with an optional 200 / 400 / 800 TOPS GPU.
10B-parameter-class large models and 128 video channels run on the device.

45TOPS

NPU compute

INT8, standard on all versions

400TOPS

GPU compute

INT8, Max

64GB

GPU VRAM

Max

128ch

1080p@30fps decoding

Plus 16-channel encoding

Versions

Four versions

Same chassis and mainboard with identical CPU, NPU, memory and storage, differentiated by GPU compute and VRAM. Standard has no GPU.

StandardCPU + NPU

45 TOPS NPU

Pro32 GB VRAM

200 TOPS GPU + 45 NPU

MaxSpec sheet model

400 TOPS GPU + 45 NPU

Pro Max128 GB VRAM

800 TOPS GPU + 45 NPU

NPU 45 TOPS (all versions) GPU (INT8)

	Standard	Pro	Max	Pro Max
Compute units	CPU + NPU	CPU + NPU + GPU	CPU + NPU + GPU	CPU + NPU + GPU
NPU (INT8)	45 TOPS	45 TOPS	45 TOPS	45 TOPS
GPU (INT8)	—	200 TOPS	400 TOPS	800 TOPS
VRAM	—	32 GB	64 GB	128 GB
CPU	Armv9 12-core	Armv9 12-core	Armv9 12-core	Armv9 12-core
Memory	16 GB LPDDR5x, up to 64 GB	16 GB LPDDR5x, up to 64 GB	16 GB LPDDR5x, up to 64 GB	16 GB LPDDR5x, up to 64 GB
Positioning	Comparable to NVIDIA Jetson Orin series	Local inference of mid-size models	Local inference of 10B-parameter-class models	Larger models and concurrent multi-task workloads

Max figures are from Product spec sheet V1.0; power consumption, dimensions and the local model list are Max values. For other versions, refer to the corresponding spec sheet.

Heterogeneous compute

Division of work across CPU, NPU and GPU

Control and scheduling run on the CPU, always-on vision and mid-size models on the NPU, and large models and parallel computing on the GPU. The three compute units work together in one device, so robots can be deployed with an integrated "large and small brain" architecture.

CPU

Armv9 12-core

A720 × 8 + A520 × 4, system > 300 KDMIPS

Task scheduling and business logic
SLAM, path planning, motion control
ROS nodes and device drivers

NPU

45 TOPS

INT8, standard on all versions

Vision detection and recognition models
Qwen3-30B-A3B, Qwen3-VL-30B-A3B
Low-power always-on inference

GPU

400 TOPS

INT8, 64 GB VRAM (Max)

Qwen3.5-122B-A10B large model
CUDA-compatible parallel computing
Radar signal and pixel-level image processing

Local large models

A 122B model,
on the device

The Max GPU runs Qwen3.5-122B-A10B; the NPU simultaneously runs 30B-class MoE language models and vision language models. Models, data and inference all stay local.

ModelTypeRuns on

Qwen3.5-122B-A10BLarge language modelGPU

Qwen3-30B-A3BLarge language modelNPU

Qwen3-VL-30B-A3BVision language modelNPU

FireRedASR2-LLMSpeech recognition—

Adapted models listed in the Max spec sheet.

Video processing

128-channel video decoding

Integrated video codec unit: 128-channel 1080p@30fps decoding and 16-channel 1080p@30fps encoding, supporting H.264, H.265 and AVS2. Decoded video streams go directly to the NPU / GPU for analysis.

16-ch encoding 128-ch decoding

Software and applications

Software ecosystem and applications

GPU ecosystem: CUDA-compatible; existing CUDA applications can be migrated
Deep learning frameworks: PyTorch, TensorFlow, ONNX Runtime, TensorRT
Multimedia frameworks: FFmpeg, GStreamer
Robotics: All ROS versions supported
Operating system: Debian 12 / Ubuntu 26.04

Robot navigation and motion control

Runs SLAM, path planning and obstacle avoidance simultaneously, working with the low-level motion control unit; the CPU handles control while the NPU / GPU run perception and decision models.

Local large model inference

Edge AI assistants, medical imaging, industrial quality inspection and similar uses, running LLMs and VLMs on the device so data stays on site.

Multi-channel video analytics

128-channel 1080p decoding with object detection, behavior recognition and anomaly detection, reducing video upload bandwidth.

Parallel computing

Data-intensive tasks such as parallel radar signal processing, large-scale pixel-level image processing and bulk data encryption.

Application

Edge–cloud hybrid for enterprise agents

AI BOX serves as the model entry point for enterprise agents. Class A deterministic tasks run as local programs, class B routine semantic tasks go to the local large model, and class C complex multi-step tasks and class D high-risk tasks call cloud models under the egress policy.

Enterprise agent

OpenAI-compatible APIIdentity · permissions · budget

ADeterministic workFormat conversion, database queries, file archiving

Local program0 tokens

BRoutine languageMeeting minutes, email summaries, extraction

Local LLM0 tokens

CComplex multi-stepMulti-document analysis, automated planning

Cloud frontier modelOnly where needed

DHigh-risk / specifiedKey decisions, high-risk code changes

Specified cloud modelOn demand

Completed locally on AI BOX Sent to the cloud under egress policy

Unified access: OpenAI-compatible API; cloud API keys are stored encrypted on the device and never issued to employees or agents
Result checks: Local output is checked for format, fields, value ranges and traceability to the source text; on failure, timeout or a complexity match, the request escalates to the cloud and the reason is logged
Egress policy: Set per data sensitivity level: no egress, egress after masking, or specified cloud vendors only
Audit and budget: Each call logs user, category, actual model, escalation reason, tokens and duration; daily / monthly budgets per user and department

Hardware

135 × 135 × 63 mm

Aluminum alloy chassis, typical power consumption < 95 W (Max). Front panel: 2 × USB 3.2 Type-C and status light; rear: DC power, 2 × RJ45, HDMI, 2 × USB 3.2 Type-A. Wi-Fi 6E, Bluetooth and 5G / 4G are optional expansions.

Technical specifications

Model	XSCD_NX_PRO (Max)
CPU	Arm A720 × 8 + A520 × 4, Armv9 architecture, 12 cores
NPU	45 TOPS @ INT8
GPU	400 TOPS @ INT8 (Pro Max: 800 TOPS)
VRAM	64 GB
Memory	16 GB LPDDR5x, up to 64 GB
Storage	512 GB NVMe SSD; TF card expansion
Network	2 × RJ45; optional Wi-Fi 6E / Bluetooth; optional 5G / 4G
Display	1 × HDMI 2.0 / 2.1, up to 4K
USB	2 × USB 3.2 Type-C, 2 × USB 3.2 Type-A
Video encoding	16 channels 1080p@30fps, H.264 / H.265 / AVS2
Video decoding	128 channels 1080p@30fps, H.264 / H.265 / AVS2
Local models	NPU: Qwen3-30B-A3B, Qwen3-VL-30B-A3B; GPU: Qwen3.5-122B-A10B; ASR: FireRedASR2-LLM
Software	Debian 12 / Ubuntu 26.04; CUDA ecosystem; PyTorch, TensorFlow, ONNX Runtime; FFmpeg, GStreamer; ROS
Power consumption	Typical < 95 W
Dimensions	135 × 135 × 63 mm
Operating temperature	0 to 50 °C (commercial grade)

Specifications from XSCD_NX_PRO Product spec sheet V1.0 (October 2026), Max configuration; the final mass-production version prevails.

AI BOX selection and testing

Tell us the model size, number of video channels, concurrent tasks and deployment environment; we will recommend a version and provide a test unit.