JWIPC W638 compact AI workstation shown in five chassis finishes

SHENZHEN, China - JWIPC has introduced the W638, a compact generative AI workstation designed for organizations and developers that need to run large language models, multimodal applications and AI agents close to their data. The system is available with NVIDIA Jetson T4000 or T5000 modules and delivers up to 2,070 sparse FP4 TFLOPS of AI performance, 128GB of unified memory and 273GB/s of memory bandwidth.

As AI workloads move from experimentation into daily operations, teams must balance model capability with cloud inference costs, latency, data governance and deployment complexity. The W638 provides a desk-side platform for workloads that can be hosted locally, allowing models and data to remain on the device while reducing dependence on metered cloud APIs.

Compute and memory for local AI

The W638 is built on the NVIDIA Jetson Thor platform and uses a Blackwell architecture GPU with fifth-generation Tensor Cores. The T5000 configuration combines a 14-core Arm Neoverse V3AE CPU with up to 2,070 sparse FP4 TFLOPS, while the T4000 configuration offers up to 1,200 sparse FP4 TFLOPS and 64GB of memory. This range lets customers choose a platform based on model size, throughput targets and deployment budget.

T5000 and T4000 configurations scale the W638 for different local AI workloads

Unified memory gives the CPU and GPU access to the same memory pool, reducing unnecessary data movement between separate memory spaces. With up to 128GB of capacity and 273GB/s of bandwidth, the W638 is designed for open-weight language models, vision-language models and long-running agent workflows. Actual model capacity and throughput depend on model architecture, quantization, context length and runtime optimization.

Local agent deployment

For workloads executed locally, the W638 can keep prompts, files, models and inference data on the workstation. This deployment model supports organizations that need tighter control over proprietary information, predictable operating costs or low-latency access without relying on a continuous cloud connection. It also removes per-request cloud API charges for inference completed on the device, although infrastructure, software and energy costs still apply.

The W638 can be supplied with NVIDIA NemoClaw or JWIPC JWiClaw software options, depending on the selected configuration. NVIDIA NemoClaw is an open-source reference stack for running supported AI agents inside NVIDIA OpenShell sandboxes with policy controls for files, networks and credentials. JWiClaw provides a visual interface intended to simplify model deployment and skill installation for users who prefer a guided workflow.

W638  can be used in many senarios

JWIPC internal performance testing

JWIPC evaluated the W638 with local inference and agent workloads. In a code-generation test on a Jetson AGX Thor configuration, the system reached 42.3 tokens per second with vLLM and 32.25 tokens per second with Ollama. A T5000 configuration with a 2TB SSD reached 90.7 tokens per second when running Qwen 30B A3B in the specified lab environment.

Performance note: Results reflect JWIPC internal testing under specific software, model and hardware settings. Performance can vary with model version, quantization, runtime, context length and system tuning.

Compact design and high speed connectivity

The W638 measures 130 x 125 x 75 mm and weighs approximately 1.3 kg, making it suitable for desk-side development, studio environments and space-constrained offices. Its air-cooled enclosure uses honeycomb ventilation to support sustained operation within a compact footprint. Storage options include 1TB or 2TB NVMe SSDs.

The W638 measures 130 x 125 x 75 mm and weighs approximately 1.3 kg

Connectivity includes two 10GbE RJ45 ports, four USB 3.2 Type-A ports, one USB Type-C port and two HDMI 2.1 outputs supporting up to 4K at 60Hz. Wi-Fi 6 and Bluetooth 5.3 are also available. The interface set supports local datasets, NAS-based workflows, multiple displays and common development peripherals without relying on additional adapters for core connectivity.

Key specifications

Product information

For specifications and configuration details visit the JWIPC W638 product page. For OEM and ODM inquiries contact overseasales@jwele.com.cn.

About JWIPC

JWIPC is a listed provider of intelligent hardware and computing solutions for AI infrastructure, industrial IoT, ICT, commercial terminals and consumer systems. The company supports customers from product definition and hardware development through system integration, validation and mass production. JWIPC is listed on the Shenzhen Stock Exchange under stock code 001339.

Technical references  JWIPC W638  |  NVIDIA Jetson Thor  |  NVIDIA NemoClaw