QAI-16700

Storage × Virtualization × On-Premises AI — the all-in-one tower Edge AI NAS

QuTS hero Operating System

Honored with 2026 MSP Today Product of the Year Award

All in One: Storage, Virtualization, and On-Premises AI

Instead of buying a NAS, a virtualization host, and a GPU AI workstation separately, the QAI-16700 combines all three roles in a single tower — greatly simplifying space, cost, and maintenance, all managed from one QuTS hero interface.

Enterprise Storage

QuTS hero ZFS with 16-bay hybrid storage plus U.2 NVMe delivers snapshots and data integrity in one package.

Virtualization Platform

Virtualization Station and Container Station, with GPU passthrough and container-native GPU access.

On-Premises AI Engine

Run LLMs and generative AI on the local GPU — data never leaves the company, covering performance, privacy, and compliance at once.

On-Premises AI, Data Stays In-House

Everything from inference to generation runs on the local GPU, so sensitive data never goes to the cloud — performance, privacy, and compliance in one.

• Private LLM Inference

  Run open-source large language models locally to build a company-specific AI assistant and Q&A service.

• Generative Imaging

  Generate images on-premises with Stable Diffusion / ComfyUI and keep every creative asset in-house.

• GPU Hardware Acceleration

  Technologies such as CUDA and TensorRT substantially raise inference throughput and response speed.

• Data Sovereignty

  All compute and data stay inside the organization, meeting privacy and compliance requirements.

AMD Ryzen™ 9 7900

A 12-core / 24-thread Zen 4 architecture combines a high core count with high clock speeds: the cores carry heavy virtualization and multitasking, while high frequencies accelerate AI pre-processing and inference scheduling. Paired with dual-channel DDR5, virtualization and on-premises AI run side by side on a single machine.

• Core Specifications

  core

  12 Cores

  thread

  24 Threads

• Max Boost Clock

  5.4 GHz

  Max Memory Capacity

  192GB DDR5

[ NAS storage illustration ]

Capacity × Speed
in One System

16 bays of hybrid storage: high-capacity HDDs as the foundation and NVMe SSDs for speed, so AI data pipelines, virtual machines, and backups each get what they need in one system.

12×

HDD · High-Capacity Storage

3.5" SATA drive arrays deliver petabyte-scale capacity — the foundation for data, backups, and snapshots.

4×

NVMe SSD · Performance Tier

U.2 / M.2 flash accelerates AI data pipelines, vector databases, and virtual machine I/O.

⇅

QuTS hero Hybrid Storage

Combine SSDs and HDDs to balance storage performance, capacity, and cost, delivering a flexible solution for diverse workloads.

Lab-Tested GPU Lineup

The 750W power supply drives a single high-end workstation GPU. Choose by model size and budget, from the 96GB flagship to a compact card; Intel Arc Pro support is coming soon.

RTX PRO 6000 Blackwell

• GPU Memory96 GB GDDR7
• ArchitectureBlackwell
• Best for70B+ LLMs · Generative AI

RTX PRO 5000 Blackwell

• GPU Memory48 GB GDDR7
• ArchitectureBlackwell
• Best for32B LLMs · Image Generation

RTX PRO 4500 Blackwell

• GPU Memory32 GB GDDR7
• ArchitectureBlackwell
• Best forMid-to-Large Model Inference

RTX PRO 4000 Blackwell

• GPU Memory24 GB GDDR7
• ArchitectureBlackwell
• Best forEntry-Level AI · Multi-Task Inference

RTX PRO 2000 Blackwell SFF

• GPU Memory16 GB GDDR7
• ArchitectureBlackwell
• Best forLightweight Inference · Compact Build

Intel Arc Pro B70

• GPU Memory32 GB GDDR6
• ArchitectureIntel Xe2 (Battlemage)
• Best forMid-to-Large Model Inference

* GPU memory and architecture are manufacturer specifications; actual support and performance depend on model configuration and firmware version.

Runs On-Premises — and Runs Fast

Measured on a single NVIDIA RTX PRO 6000 Blackwell Max-Q: large language models from 4B to 120B parameters all run inference at high speed on-premises.

0tok/s

Peak Total Throughput · vLLM, 8 Streams

0tok/s

Single Stream · Ollama, One-User Feel

0B

Parameter Model on a Single Card

Ollama single stream (one-user feel)vLLM 8 streams (peak throughput)

gpt-oss 120B

106 / 1,047

Gemma 4 26B-A4B

192 / 998

Gemma 4 E4B

177 / 855

Qwen 27B

64 / 354

Measured on an NVIDIA RTX PRO 6000 Blackwell Max-Q 96GB (512-token prompt / 256-token generation, temperature 0); the QAI-16700 ships with the same GPU. Single-stream figures use the native Ollama API (pure GPU decode rate); peak figures use vLLM /v1 with 8 concurrent streams.

Click to Feel the Generation Speed

Tokens per second decide how fast an AI reply “types.” Pick any measured model speed below, watch the rhythm it generates at in a chat interface, and compare it with a cloud API.

gpt-oss 120B0 tok/s0 tokens

Introduce the QAI-16700 in one sentence

Q

← Pick a button below to start generating

Try different speeds:

Typing pace is simulated from measured single-stream tok/s; the text is illustrative.

Qsirch & RAG Search

Turn Company Documents into a Private, Conversational Knowledge Base

Qsirch full-text search combined with RAG turns internal documents into a private, queryable knowledge base linked to a local LLM — search and generation both stay on-premises, so knowledge never leaks out.

Learn More: Learn about Qsirch

A More Powerful Container Station for One-Stop On-Premises AI

Container Station on the QAI-16700 includes a rich catalog of AI app templates for one-click deployment of mainstream AI tools and frameworks, continuously updated by QNAP. Whether you are new to AI or moving existing workloads back on-premises, you can adopt AI easily, cut costs, strengthen data security, and even build your own AI tools to drive the business forward.

Container Station template screen (appears once container-templates.png is uploaded)

One-Click On-Premises AI Solutions

From inference engines and chat interfaces to vector search and RAG knowledge bases, every solution below runs locally through Container Station with GPU integration — your data never leaves the company.

Ollama

Run and experiment with open-source language models the easy way: GPU acceleration lets developers test, fine-tune, and serve LLMs efficiently on-premises, with data staying local.

Open WebUI

A full-featured, easy-to-use web interface for interacting with local or remote LLMs, with chat, prompt engineering, and conversation management that make AI adoption simpler.

AnythingLLM

Deploy a customizable private large language model in a secure environment: connect multiple data sources for tailored Q&A and document analysis — ideal for businesses that value privacy and control.

RAGFlow

A document RAG engine: a built-in vector database and document chunking turn contracts, specifications, and presentations into a searchable vector index, so a local LLM answers from your actual documents and cites its sources.

Dify

An LLM application development platform: a built-in vector knowledge base and prompt-flow orchestration let you build a company-specific Q&A bot without coding from scratch, then expose it to other systems through an API.

Flowise

A drag-and-drop AI flow builder: chain large language models, vector search, and external APIs into agents or Q&A flows, so teams that do not code can assemble their own AI applications.

ComfyUI

An advanced modular generative AI workflow interface: build complex AI pipelines such as image generation by dragging nodes, all GPU-accelerated.

Whisper

A single open-source model for multilingual speech recognition: accurate transcription, translation, and language identification, deployable on-premises for both data security and voice application integration.

Tabby

An on-premises AI coding assistant: code completion and Q&A on your own GPU, with source code that never leaves the company — ideal for development teams with compliance requirements.

Real-World AI Scenarios, from Enterprises to Studios

A single QAI-16700 supports a wide range of on-premises AI scenarios.

• Enterprise AI Assistant

  Build a private Q&A assistant on internal documents so employees get instant answers on policies and data.

• Media and Design Studios

  Collaborative NAS storage alongside ComfyUI / SD local image generation and editing VMs keeps every asset in-house.

• Software Development Teams

  A local code assistant, development VMs, and version control / CI storage, all in one place.

• Process Automation

  Connect a local LLM through n8n to automate document handling, notifications, and data organization.

Real ComfyUI, Running on the Local GPU

This is a ComfyUI node workflow running on the QAI-16700. Press Queue Prompt and watch it run node by node until Save Image produces the picture — entirely on-premises, with assets that never leave the company.

🔒 qai-16700.local:8188  ·  ComfyUI

Ready · 7 nodes

Queue Prompt

← Scroll sideways to see the full workflow →

Your Own ChatGPT, Running Inside the NAS

A familiar chat interface — but the large language model behind it runs entirely on the QAI-16700's local GPU. Try a question: not one byte of data leaves your server room.

🔒 qai-16700.local  ·  Private LLM · 100% on-device

QAI-16700 · Llama 3.1 70B ▾Running locally

Q

Hi! I'm an AI assistant running entirely on this QAI-16700. Ask me anything — your conversation stays on this machine.

👇Pick any question below and watch it answered live on the local GPU

↑

QAI-16700 local model · no cloud connection · conversations stay in-house

Why Choose the QAI-16700

Compared with other AI NAS brands, GB10-class mini AI appliances, or large GPU servers, the QAI-16700 brings enterprise storage, virtualization, and on-premises GPU AI together in a single deskside tower.

	QAI-16700	Other AI NAS brands	GB10 mini AI appliance	Large GPU server
On-premises GPU AI inference	✔	Partial	✔	✔
Swappable high-end workstation GPU (RTX PRO 6000 class)	✔	✘	GPU soldered	✔
Enterprise ZFS storage + snapshots	✔	Partial	✘	Partial
High-capacity drive bays	16-bay	4–8	No built-in array	Sold separately
Virtualization (VM / container)	✔	Partial	✘	✔
Unified management interface (QuTS hero)	✔	Partial	✘	✘
Deskside deployment · no server room	✔	✔	✔	✘

* GB10 refers to NVIDIA Grace Blackwell-class mini AI computers (such as DGX Spark-style appliances). Some items depend on configuration; this comparison is for positioning reference only.

One QNAP Edge AI Family — Choose by Scale

From a deskside entry point to a data-center flagship, a complete on-premises AI upgrade path.

QAI-16700 product image

Deskside · High Capacity

QAI-16700

The on-premises AI starting point for SMBs, studios, and R&D teams: a tower chassis, Ryzen 9 plus a single high-end GPU, and 16-bay hybrid storage — deployable right in the office.

Data Center · Flagship

QAI-h1290FX

A flagship platform for larger scale and higher throughput: server-grade processors, U.2 all-flash storage, and higher GPU / memory ceilings — the upgrade of choice as compute demand grows.

QAI-16700 on the Show Floor

Visiting trade shows and customer sites worldwide, so you can experience the QAI-16700's on-premises AI firsthand.

QuTS hero · ZFS

Enterprise ZFS That Safeguards Your AI Assets

QuTS hero, built on ZFS, delivers self-healing, near-unlimited snapshots, inline deduplication, and SSD endurance optimization. Keep model weights, training data, and vector databases on storage with end-to-end integrity protection, and your AI results stay safe.

Learn More: Learn about QuTS hero

Virtualization, Ready to Work

GPU passthrough and container-native GPU access turn one machine into a flexible virtualization and AI platform.

• Virtualization Station

  Create Windows / Linux virtual machines with GPU passthrough for AI or GPU workstation applications.

• Container Station

  Native GPU access inside containers, with one-click deployment of tools such as Ollama, OpenWebUI, and ComfyUI.

• VM High Availability

  Clustered deployment and failover raise the availability of critical virtual machines and services.

• Virtual Switch

  Flexibly configure virtual networks to connect multiple VMs and container services.

FAQ

1

Who is the QAI-16700 for?

It is built for SMBs, creative studios, and R&D teams that need to run AI locally while also covering storage and virtualization. One system integrates enterprise ZFS storage, a virtualization platform, and GPU AI inference — no cloud, and no fleet of machines to manage.

2

Which graphics cards are supported?

With a 750W power supply it accepts high-end discrete graphics cards; the NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation GPU is supported in lab testing. See the hardware specification page for the full compatibility list.

3

Can it run virtual machines and AI at the same time?

Yes. The Ryzen 9 7900 provides 12 cores and 24 threads, and with GPU passthrough and container-native GPU access it supports virtualization and AI inference workloads at the same time.

4

Does AI inference data go to the cloud?

No. All inference and generation run on the local GPU, and data stays entirely inside your organization.

5

Which AI applications can be deployed?

Container Station deploys mainstream open-source tools such as Ollama, OpenWebUI, AnythingLLM, ComfyUI, n8n, and Whisper with a single click.