View 01 · Front View with Industrial Perforated Air Intake

AI WORKSTATION · LOCAL COMPUTE WORKSTATION

AiW-V Series

Accessible 32GB/64GB industrial VRAM workstation bringing mainstream LLMs and offline AI video creation to your desk.

Purpose-built for budget-conscious innovators. Powered by 32GB/64GB HBM2 VRAM, running Qwen 27B-class LLMs and ComfyUI offline pipelines with zero cloud token anxiety and 100% local data privacy.

32GB / 64GB HBM2 VRAMVerified Qwen 27B FluencyComfyUI Offline Video Pipeline100% On-Site Data Privacy

Request a Quote →Compare Models ↓

CORE VALUES

Take the First Step in Local AI Without the Heavy Cost.

Direct access to 32GB/64GB industrial physical VRAM at an accessible entry point.

01 / VRAM CAPACITY

32GB / 64GB Industrial VRAM

Overcome consumer GPU memory bottlenecks. 32GB or 64GB HBM2 VRAM with 900 GB/s bandwidth comfortably fits 27B-class LLMs and diffusion video pipelines without OOM errors.

02 / ZERO SUBSCRIPTION

Zero Token Billing Anxiety

Developing and experimenting with AI requires constant iteration. A local workstation eliminates recurring per-token fees, providing unlimited 24/7 experimentation with zero cost anxiety.

03 / DATA PRIVACY

100% Data Kept On-Premises

Internal SOPs, unreleased client briefs, and confidential codebase stay strictly within your local network. Physical boundary isolation prevents proprietary data leaks.

REAL BENCHMARK TELEMETRY

Qwen3.8-27B Local Benchmark: Verified Telemetry Data.

Transparent engineering telemetry. Verified in AiW-V environments with optimized vLLM containers, demonstrating stable and resilient local inference.

▲ Real performance telemetry: Dual V100 running Qwen3.8-27B under vLLM + MTP speculation, showing perfect concordance between concurrent throughput and token counters.

131,072 tokens

128K Max Context Stress Tested

Sustains 128K ultra-long context stress testing; long manuals, codebases, and extensive documents process smoothly without memory exhaustion.

425.67 tok/s

Peak Concurrent Aggregate Throughput

Under 8 concurrent requests, MTP4 speculative decoding delivers up to 425.67 tok/s aggregate throughput, powering seamless multi-workstation access.

OpenAI API

Native Standard API Interface

Exposes standard OpenAI-compatible API out of the box, integrating with Dify, FastGPT, and team clients in under 5 minutes.

Configuration & Conditions	Concurrency	Tokens / Run	Total Time	Actual Throughput	Scheduling State
AiW-V32 Single 32GB (Quantized)	1 Stream	256 tokens	~6.98s	~36.7 tok/s	Smooth for individual workflows
AiW-V64 Dual GPU (32K Ctx + MTP4)	8 Streams	7,200 tokens (900/路)	16.91 s	425.67 tok/s	8 concurrent requests zero queue
AiW-V64 Dual GPU (128K Stress + MTP4)	3 Streams	6,000 tokens (2000/路)	25.28 s	237.26 tok/s	Highly stable long-context generation

CREATIVE PIPELINE

ComfyUI Offline Video Studio: Offload Waiting to Local Compute.

Video diffusion models demand high physical memory. With 32GB+ VRAM, AiW-V offers creators a dedicated offline rendering station with zero per-run fees.

Image-to-Video Motion Showcase

Converts static white-background product shots or sketches into fluid 3-5s clips with 3D orbiting camera moves for e-commerce and social feeds.

E-Commerce Showcase3D Orbiting Motion

Overnight Batch Rendering

Queue complex prompt pipelines before leaving the office; let the station render silently overnight and collect high-res results in the morning with zero bill shock.

Long-Run BatchingZero Token Cost

Digital Avatar Video Base Plates

Renders fluid portrait motion plates for lip-sync and virtual avatar narration, significantly lowering the overhead of video production.

Avatar Base PlatesNatural Facial Motion

PRACTICAL USE CASES

Practical Workflows: 4 Core Scenarios for Small Teams.

Focused on everyday high-frequency workflows for growing teams and creators.

APPLICATION AREA 01 · ENTERPRISE KNOWLEDGE BASE

Private Team Knowledge Base & Document QA

Workflow Need: Teams need AI-powered search across internal SOPs, product specs, and sales decks on a budget, without uploading proprietary data to cloud LLMs.

AiW-V Solution: Hosts open-source RAG tools paired with a 27B-class LLM over LAN. Colleague browsers access instant answers with zero data egress outside the office.

27B LLMStable local model inference

100% LANConfidential data stays on-prem

Note: Model inference runs locally on AiW-V hardware. RAG frontends and vector databases deploy on-chassis or across the LAN.

Private Team Knowledge Base & Document QA

CHASSIS & HARDWARE

Cold-Rolled Steel Chassis & Direct Through-Flow Cooling.

All-metal reinforced body engineered for 24/7 compute. Top and side ventilation channels sustain optimal thermals and prevent throttling.

1× or 2× V100 32GB

Industrial Compute Units

Equipped with 1 or 2 NVIDIA V100 32GB industrial GPUs, delivering steady horsepower for LLMs and diffusion.

32GB / 64GB HBM2

High-Bandwidth VRAM

32GB or 64GB HBM2 VRAM boasting 900 GB/s memory bandwidth, unlocking smooth local token throughput.

Intel Xeon E5-2696

Multi-Core Enterprise CPU

Proven multi-core enterprise server CPU on X99 architecture, ensuring solid OS scheduling and I/O bus throughput.

32G / 64G + 512G SSD

System RAM & NVMe SSD

Standard 32GB/64GB DDR4 system RAM and 512GB fast NVMe SSD for swift model weight loading and caching.

Direct Airflow Chassis

Heavy-Duty Thermal Design

Rigid cold-rolled steel enclosure with heavy perforated mesh panels, generating direct airflow to prevent thermal throttling.

750W / 1000W Power

Generous Power Headroom

Fitted with rated 750W (V32) / 1000W (V64) power supplies for sustained multi-day compute sessions.

1 × Gigabit RJ45

Gigabit LAN Connectivity

Standard Gigabit Ethernet port connects directly to office switches, serving local PCs and mobile devices instantly.

440×395×165 mm

Desk-Ready Compact Form

Compact chassis (14.9~15.8 kg) tucks neatly beside or beneath standard desks, requiring no server rack or room.

SPECIFICATIONS

AiW-V32 vs. AiW-V64 Model Comparison.

Unified industrial steel enclosure across both models. Choose 32GB single GPU or 64GB dual GPU based on workload and team size.

Specification	AiW-V32 Entry · Solo Creators	AiW-V64 Advanced · Teams & Multi-Task
GPU Configuration	1 × NVIDIA V100 32GB	2 × NVIDIA V100 32GB
Total VRAM	32GB HBM2 VRAM	64GB HBM2 (32GB × 2)
Memory Bandwidth	Up to 900 GB/s	Up to 900 GB/s (per card)
Target Models & Workloads	Qwen 27B Quantized / ComfyUI single-stream	Qwen3.8-27B 128K context / Concurrent streams
Processor	Intel Xeon E5-2696 Enterprise Multi-Core	Intel Xeon E5-2696 Enterprise Multi-Core
System Memory	32GB DDR4 RAM	64GB DDR4 RAM
Local Storage	512GB NVMe SSD SSD	512GB NVMe SSD SSD
Power Supply	750W Rated Power Supply	1000W Redundant Power Supply
Cooling Architecture	Perforated direct airflow + PWM fans	Perforated direct airflow + PWM fans
Network Interface	1 × 千兆 RJ45 (GbE RJ45 pure wired)	1 × 千兆 RJ45 (GbE RJ45 pure wired)
Operating System	Ubuntu Linux pre-installed (vLLM & ComfyUI)	Ubuntu Linux pre-installed (vLLM & ComfyUI)
Dimensions & Weight	440 × 395 × 165 mm · ~14.9 kg	440 × 395 × 165 mm · ~15.8 kg
Recommended Audience	Solo creators, vloggers, university students	SMB teams, multi-seat LAN, university labs

Selection Note: Both models share the industrial cold-rolled steel enclosure and desk-ready footprint. AiW-V32 covers single-stream workloads; choose AiW-V64 for multi-seat LAN concurrency, 128K long-context stress loads, or concurrent LLM + video diffusion tasks.

FULL-STACK ON-SITE AI SOLUTION

On-Site AI Synergy: AiR + AiA + AiW Triad.

From secure network ingress to 24/7 resident agents and local LLM compute. Three distinct appliances form a sovereign on-site AI stack.

STEP 01 / NETWORK

AiR AI Cloud Gateway

Network & Boundary: Provides clean LAN infrastructure, outbound proxy isolation for external APIs, and reverse tunneling for remote access at 12W power draw.

Explore AiR Gateway →

STEP 02 / AGENT RUNTIME

AiA AI Agent Host

24/7 Agent Runtime: Decouples agents from personal laptops. Runs 24/7 multi-agent workflows, managing business orchestration while dispatching compute jobs to AiW.

Explore AiA Host →

STEP 03 / COMPUTE BRAIN

AiW-V Local Compute Brain

Core Compute Brain: Provides 32GB/64GB HBM2 VRAM, running local Qwen 27B models and ComfyUI pipelines over standard APIs with 100% on-site data custody.

Current Model · AiW-V Details →

FAQ

Frequently Asked Questions.

Clear answers on memory architecture, model selection, API integration, and creative workflows.

For LLMs and video diffusion, physical VRAM capacity is the fundamental ceiling. Consumer gaming cards (16G/24G) frequently crash with Out of Memory errors when loading 27B-class models with long context or multi-step diffusion pipelines. Multi-card consumer setups also suffer severe thermal throttling in desktop cases. AiW-V delivers 32GB (single GPU) or 64GB (dual GPU) of industrial HBM2 memory (900 GB/s bandwidth) in a custom ventilated steel chassis, running mainstream models at a fraction of the cost of flagship enterprise servers.

Build Budget-Friendly Local AI Compute for Your Team

Say goodbye to recurring cloud token bills and unlock 32GB/64GB industrial VRAM today.

Discuss Product Selection →