Putting Large Models
into a Small Box
——AIBOX-1684X AI Box
Up to 32 TOPS of computing power! The AI box supports the private
deployment of mainstream large models,
bringing private AI capability to meet individual AI deployment needs
The private deployment of complete
large language models

32TOPS BM1684X AI processor
with ultra-high computing power

BM1684X, the SOPHON AI processor, features an octa-core ARM Cortex-A53 with up to 2.3 GHz of frequency and a 12nm lithography process.
With up to 32Tops (INT8) computing power, 16TFLOPS (FP16/BF16), or 2Tops (FP32) high-precision computing power,
it supports mainstream programming frameworks, and can be widely used in artificial intelligence inference for cloud and edge applications.

Powerful multi-channel video AI performance

The AI box supports up to 32 channels of 1080P H.264/H.265 video decoding and 32 channels of 1080P HD video processing (decoding +
AI analysis),making it ideal for various AI applications such as face detection and license plate recognition on video streaming.

Strong network communication capability

With dual 1000Mbps Ethernet, the AI box ensures high-speed and stable network communication,
meeting the needs of various application scenarios.

All-aluminum alloy enclosure for heat dissipation

The industrial-grade all-metal enclosure with aluminum alloy structure for thermal conduction. The side of the top cover features a
grille design for external airflow and efficient heat dissipation, ensuring computing performance and stability even under
high-temperature operating conditions. Its top cover is a porous hexagonal design, combining elegance with high efficiency.
The compact, exquisite device operates stably and meets the needs of various industrial-grade applications.

Abundant resources

We offer SDKs, tutorials, technical documentation, and development tools to streamline and improve the development process

A wide range of applications

The device is widely used in intelligent surveillance, AI education, services based on computing power, edge computing,
private deployment of large models, and data security and privacy protection

AI education
Computing services
Edge computing
Private deployment
of large models
Data security


Basic Specifications SOC


High-performance octa-core ARM A53, 12nm lithography process, frequency up to 2.3 GHz


Built-in tensor computing module TPU, computing power up to: 32TOPS (INT8), 16TFLOPS (FP16/BF16), 2TFLOPS (FP32)


32-channel H.265/H.264 1080p@25fps video decoding,

32-channel 1080P@25fps HD video processing (decoding +AI analysis),

12-channel H.265/H.264 1080p@25fps video encoding




32GB/64GB/128GB eMMC、1*TF Card


DC 12V/4A (DC 5.5*2.5mm)




・ The private deployment of ultra-large-scale parameter models under the Transformer architecture,

including large language models such as LLaMa2, ChatGLM, and Qwen, as well as large vision models like ViT,

Grounding DINO, and SAM.

・ The private deployment of the Stable Diffusion V1.5 image generation model in the AIGC field.

・ Traditional network architectures such as CNN, RNN, and LSTM; a variety of deep learning frameworks,

including TensorFlow, PyTorch, MXNet, PaddlePaddle, ONNX, and Darknet as well as custom operator development

・ Docker container management technology


90.6mm * 84.4mm * 48.5mm


≈ 420g


Operating: -20℃~60℃, storage: -20℃~70℃, humidity:10%~90%RH (non-condensing)

Interfaces Ethernet



2*USB3.0(current limit 1A)


1*power button, 1*Type-C (Debug serial)