How to Run Ollama on a QNAP NAS — Setup Guide 2026

QNAP NAS units with AMD Ryzen processors are the strongest CPU-based local LLM hardware in AU retail, and Ollama via Container Station makes deployment straightforward. The TS-473A (AMD Ryzen V1500B, up to 64 GB RAM) outperforms equivalent Synology hardware for LLM inference because it includes AVX2 SIMD support and a 4-core rather than 2-core processor. This guide covers QNAP-specific setup from hardware check through to a fully functional Ollama instance with Open WebUI, plus the performance expectations and configuration details that differ from the Synology setup. For the Synology equivalent, see Ollama on Synology NAS.

Compatible QNAP Models for Ollama

Ollama requires an x86 processor on QNAP hardware. ARM-based QNAP units cannot run it. Most current mainstream QNAP models are x86, but performance varies significantly by CPU generation and AVX2 support.

Recommended QNAP models (AU retail, best to adequate):

• TS-473A (from $1,269 at Scorptec): AMD Ryzen V1500B, 4-core/8-thread, AVX2. Four RAM slots, supports 64 GB. The top recommendation for LLM on NAS in AU retail. Upgrade to 16 GB RAM before running 7B models.
• TS-873A (from $1,829 at Scorptec/Computer Alliance): Same AMD Ryzen V1500B CPU, 8-bay, two PCIe slots. Better for users who also need 10GbE alongside AI (can use one slot for each), or for future GPU addition.
• TS-673A (from $1,599): 6-bay, same AMD platform. Upgrade from TS-473A if you need more drive bays.
• TS-664 (from $1,249): Intel Celeron N5105, 6-bay. No AVX2, slower inference, but workable for 3B models. Better for photo AI and OCR than LLM.
• TS-464 (from $989): Intel Celeron N5105, 4-bay. Entry-level option, no AVX2, limited to 8 GB max RAM. Runs 3B models at adequate speed.

Models to avoid for LLM: TS-133, TS-233, and any QNAP model with ARM architecture. Check the QNAP product spec page; "Cortex" in the CPU field = ARM = Ollama not supported.

Step 1: Install Container Station

Container Station is QNAP's Docker runtime, available through the App Center.

1. Log into the QNAP QTS web interface.
2. Open the App Center.
3. Search for Container Station.
4. Click Install. The application is free.
5. Once installed, open Container Station from the QTS main menu.

Container Station on QTS uses a slightly different interface from Docker Desktop, but the underlying engine is standard Docker CE. Containers, images, and Docker Compose files work identically to any other Docker host.

Step 2: Create Shared Folders for Container Data

Create dedicated folders for Ollama model storage and Open WebUI data before deploying containers.

1. Open File Station in QTS.
2. Navigate to your main data volume (typically DataVol1).
3. Create a folder named docker.
4. Inside docker, create two subfolders: ollama and open-webui.

If your QNAP has an NVMe M.2 SSD installed and configured as a cache or storage volume, storing the ollama folder there significantly reduces model load times. A 7B model stored on NVMe loads into RAM in approximately 5-10 seconds. The same model on a spinning HDD takes 60-120 seconds. The NVMe does not affect inference speed once the model is in RAM.

Step 3: Deploy Ollama and Open WebUI via Docker Compose

Container Station supports Docker Compose projects. This is the cleanest deployment method.

1. Open Container Station.
2. Click Applications in the left sidebar (the Docker Compose interface).
3. Click Create.
4. Name the application ollama.
5. Paste the following Docker Compose configuration:

version: '3.8'
services:
  ollama:
    image: ollama/ollama:latest
    container_name: ollama
    restart: unless-stopped
    ports:
      - "11434:11434"
    volumes:
      - /share/DataVol1/docker/ollama:/root/.ollama
    environment:
      - OLLAMA_HOST=0.0.0.0
      - OLLAMA_NUM_PARALLEL=1
      - OLLAMA_MAX_LOADED_MODELS=1

  open-webui:
    image: ghcr.io/open-webui/open-webui:main
    container_name: open-webui
    restart: unless-stopped
    ports:
      - "3000:8080"
    volumes:
      - /share/DataVol1/docker/open-webui:/app/backend/data
    environment:
      - OLLAMA_BASE_URL=http://ollama:11434
    depends_on:
      - ollama

The QNAP shared folder path format is /share/DataVol1/ (not /volume1/ as on Synology). Adjust the path if your volume is named differently. OLLAMA_MAX_LOADED_MODELS=1 ensures only one model is loaded at a time, which is important on NAS hardware with limited RAM.

Click Validate and then Create. Container Station pulls the images and starts both containers. Image pull takes 3-10 minutes depending on connection speed.

Step 4: Pull AI Models

Use Container Station's terminal to pull models into the Ollama container.

1. In Container Station, click Containers.
2. Find the ollama container and click it.
3. Click the Terminal icon (command prompt).
4. In the terminal, run: ollama pull llama3.2:3b

Alternatively, use SSH:

1. SSH into the QNAP: ssh admin@your-qnap-ip
2. Run: docker exec -it ollama ollama pull llama3.2:3b

Recommended models for QNAP hardware (by use case):

• llama3.2:3b. Fast, low RAM, good general assistant. Start here.
• llama3.2:7b. Better quality, needs 16 GB RAM total. Best general-purpose model for TS-473A with upgraded RAM.
• mistral:7b. Strong for structured tasks, coding, and analysis. Same RAM requirement as Llama 7B.
• qwen2.5-coder:7b. Best for code review and generation at the 7B tier.
• phi3:mini. 3.8B, very efficient for its size, fast response on QNAP hardware.

Run ollama list in the container terminal to see all installed models and their sizes.

Step 5: Access Open WebUI

Open a browser on any device on the same local network. Navigate to http://your-qnap-ip:3000. Open WebUI loads and prompts for account creation on first visit. Create an admin account (this is local only, not connected to any external service).

Select a model from the dropdown at the top of the chat window. Send a message. The first response after selecting a model may take 10-30 seconds as the model loads into RAM. Subsequent messages in the same session are faster as the model remains loaded.

Open WebUI supports multiple conversations, file uploads (PDF and text for document context), system prompt configuration, and model switching. It is compatible with mobile browsers, making it accessible from phones and tablets on the same Wi-Fi network.

QNAP Performance vs Other Options

The AMD Ryzen V1500B in the TS-473A delivers the best CPU-only LLM performance available in a standard 4-bay AU NAS at this price point.

TS-473A with 16 GB RAM:

• Llama 3.2 3B: approximately 8-12 tokens per second
• Llama 3.2 7B: approximately 4-7 tokens per second
• Mistral 7B: approximately 4-6 tokens per second

TS-464 with 8 GB RAM (Intel N5105, no AVX2):

• Llama 3.2 3B: approximately 4-6 tokens per second
• Llama 3.2 7B: marginally functional, 2-3 tokens per second with memory pressure

The AVX2 advantage of the AMD platform is approximately 1.5-2x in token generation speed for the same model size. For users who use Ollama regularly, this is a meaningful quality-of-life difference in response latency. For background batch processing (where latency does not matter), either platform is functional.

GPU acceleration option: The TS-473A's PCIe x4 slot supports a low-profile or standard-height GPU (depending on case). Adding an NVIDIA RTX 3060 12 GB (approximately$1269-$1818 AU) upgrades inference from 4-7 tokens per second (CPU) to 40-70+ tokens per second (GPU). This requires configuring Ollama with NVIDIA GPU support (CUDA), which involves additional container configuration. For hardware requirements details, see AI NAS Hardware Requirements.

For running cost modelling of any QNAP AI NAS configuration, use the NAS Power Cost Calculator. For a broader comparison of NAS AI hardware options, see Best NAS for AI Australia. For NAS selection advice outside of AI use cases, see Best NAS Australia.

Australian Buyers: What You Need to Know

AU retail sourcing. The QNAP TS-473A is available from Scorptec, PLE Computers, Mwave, Computer Alliance, and Computer & Parts Land, typically from $1,269. RAM upgrades (DDR4 SO-DIMM, standard non-ECC) are available from Scorptec, MSY, and Mwave. For the TS-473A specifically, a 12 GB or 16 GB SO-DIMM adds approximately $60-100 from AU retailers and brings total RAM to 16 GB or 20 GB respectively.

Privacy and data sovereignty. All inference runs on the QNAP hardware in your location. Prompts, documents, and responses never reach any external server. For businesses that need to process client data, medical records, or commercially sensitive content, local NAS inference provides clear data sovereignty under Australian jurisdiction, avoiding the cross-border disclosure obligations under the Privacy Act 1988 that cloud AI services can trigger.

QNAP support for Container Station. Container Station is a supported QNAP product. Docker containers running within it are a supported use case. Ollama itself is not a QNAP-tested application, but it runs on standard Docker infrastructure that QNAP supports. If you contact QNAP support about Container Station issues unrelated to Ollama, the unsupported container running alongside should not affect your support eligibility for the Container Station platform itself.

AU electricity cost of QNAP AI inference. The TS-473A draws approximately 25-40W under CPU AI inference. At AU electricity rates of 30-35 cents per kWh, continuous inference (not typical for most use cases) costs approximately $65-100 per year in additional power beyond storage-only idle costs. Model the specific cost for your usage pattern at the NAS Power Cost Calculator. Australian Consumer Law protections apply to QNAP hardware purchased from AU retailers.

Related reading: our Synology vs QNAP comparison, our NAS vs cloud storage comparison, and our NAS explainer.

Free tools: NAS Sizing Wizard and AI Hardware Requirements Calculator. No signup required.

Which QNAP NAS is best for Ollama in Australia?

The QNAP TS-473A is the best AU-available option for Ollama. Its AMD Ryzen V1500B with AVX2 produces 4-7 tokens per second for 7B models, approximately twice the speed of Intel Celeron alternatives. It supports up to 64 GB RAM across four slots. Upgrade to 16 GB total before running 7B models. Available from $1,269 at Scorptec, Mwave, and other AU retailers. For 8-bay storage alongside AI, the TS-873A (same CPU platform) is the step up.

Can I use Ollama on QNAP without Container Station?

Container Station is the standard and easiest path. It is possible to install Docker CE directly on QNAP via the QTS Linux environment, but this is significantly more complex and not recommended. Container Station provides a supported Docker environment with GUI management, image updates, and log access. Using Container Station is the right approach for all standard Ollama deployments.

Can I add a GPU to my QNAP for faster Ollama inference?

Yes, on QNAP models with PCIe expansion. The TS-473A has one PCIe 2.0 x4 slot. An NVIDIA RTX 3060 12 GB fits (low-profile bracket required for some cases) and upgrades 7B model inference from 4-7 tokens per second to 40-70+ tokens per second. GPU support in Ollama requires the NVIDIA container toolkit and a modified container configuration. The TS-873A has two PCIe slots, allowing a GPU and a 10GbE card simultaneously. For full details on GPU compatibility and power requirements, see AI NAS Hardware Requirements.

How is QNAP Container Station different from Synology Container Manager?

Both are Docker CE runtimes with web-based GUI management. The practical differences: Container Station on QTS has a slightly more complex UI with more options exposed. Container Manager on DSM has a cleaner, more streamlined interface. Docker Compose works the same on both platforms. The main path differences are folder paths (QNAP uses /share/DataVol1/, Synology uses /volume1/) and the shared folder creation workflow. Performance differences come from the underlying CPU hardware, not the Docker runtime.

Will Ollama containers survive a QNAP restart?

Yes, if configured with restart: unless-stopped in the Docker Compose file (included in the configuration above). After a QTS reboot, Container Station starts automatically, and the Ollama and Open WebUI containers restart automatically. The first model response after a reboot will include model loading time (5-30 seconds depending on whether NVMe or HDD storage is used for model weights).