Synology NAS and Ollama: local AI on your storage

By Emmanuel Sauvage · Published · Updated

This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.

You already have a Synology for family backups and photos. Adding Ollama there is tempting: one box already on, and no new computer in the living room. DSM (DiskStation Manager, the interface and operating system of Synology NAS units) 7.x is used here through Container Manager (Docker). It works for small language models and internal APIs (home automation, scripts), if you accept higher latency than a mini PC (a small desktop, often Ryzen or an Intel NUC) or a Mac mini (a compact Apple computer with unified memory). This guide covers volumes, an API test with curl (a command-line tool), and an honest comparison of a NAS and a mini PC.

About 20 minutes of reading. Before you open Container Manager, read the common questions (DS218+ compatibility, RAM, GPU, Docker, guest Wi-Fi, ChatGPT). They sit in one block at the end of the page, and they do not repeat the steps. Then follow the path below.

Profile Path Rough time What you need first
Beginner (first NAS plus Docker) Common questions → What you need → steps 1 to 5 (install) ~20 min of reading (install time is extra) An x86_64 NAS, DSM 7.2+, and either a first local AI in 15 minutes or Ollama already seen on a PC
Intermediate (NAS plus mini PC) Latency comparison plus uses that hold up +15 min An Ollama container that already runs, or a quick read of steps 1–5

Guide contents

What a NAS does well for AI, and where it falls short

Strengths: it is already on 24 hours a day, it is a central store for GGUF files (the quantized model format for Ollama and llama.cpp), and it sits naturally next to Home Assistant (open-source home automation) or n8n (workflow automation between apps) on the same LAN (the home network). Overall power use stays moderate when inference is rare.

Limits: consumer models have almost no dedicated GPU (graphics processor). An embedded Atom, Celeron, or Ryzen is limited in AVX (the CPU instructions these models use) and in sustained clock speed. RAM is shared with DSM, Synology Photos, other media services (DSM 7), Docker, and sometimes Virtual Machine Manager. Large 13B+ models, even with aggressive quantization, stay slow and can force swap (the system moves part of RAM onto the disk, which hurts latency badly) when memory is tight.

A realistic NAS-only job: short summaries, text classification, one-sentence home-automation replies, 3B–7B Q4 models. A poor fit: a smooth family assistant, heavy Continue (the editor extension for AI coding), images without a cloud, or ComfyUI and FLUX on the NAS.

What you need

Item Recommendation
Model A Synology x86_64 (a recent Plus series, or higher)
RAM 16 GB for a comfortable minimum; 32 GB if you run several containers
Model storage A folder on an SSD volume or an SSD cache, so reads stay quick
DSM 7.2+ with Container Manager (the app formerly called Docker)
Network Gigabit Ethernet. Skip Wi-Fi for large model downloads (the Ollama pull command, from the registry)
Backup A snapshot of the Docker volume before major updates

Check the maximum RAM for your chassis (often a SODIMM, a laptop-style memory stick, bought separately). Cross-check 32 vs 64 GB of RAM if you are choosing between a NAS upgrade and a new mini PC.

Install Ollama on DSM (Container Manager)

Menu names shift a little between DSM versions. The logic stays the same.

1. Prepare the folders on the NAS

In Control Panel → Shared Folder, create for example:

  • docker/ollama for local configuration, if you need it;
  • ollama-models for the model weights (this folder can grow by tens of gigabytes).

Put ollama-models on a Btrfs volume with checksums, on an SSD if you can (SSD storage for models).

2. Install Container Manager

Package Center → Container Manager. On first launch, accept the network permissions. You do not need Kubernetes for Ollama alone.

3. Create the Ollama container

Container → Create → Create with Docker Hub (or import a compose file).

The usual official image is ollama/ollama:latest. On the Docker registry, confirm the linux/amd64 variant for an Intel or AMD Synology.

Useful environment variables:

  • OLLAMA_HOST=0.0.0.0 so the container listens on all of its interfaces. You need this for access from another device on the LAN.

Volumes (mounts):

Path in the container Path on the NAS Role
/root/.ollama docker/ollama or a subfolder Ollama metadata
/root/.ollama/models ollama-models Downloaded models (they persist)

Alternative: one mount of /root/.ollama to a parent folder that holds everything.

Port:

  • Publish 11434 (TCP): container 11434 → NAS 11434 (or another host port if 11434 is taken).

Resources:

  • CPU limit: leave at least 2 cores for Ollama. Leave DSM enough CPU for itself.
  • RAM limit: for example 6–12 GB, depending on the NAS total (leave 4 GB or more for the system).

Device access: on many models there is nothing to enable for a GPU. Inference stays entirely on the CPU, with no graphics acceleration.

Start the container and read the logs. The server should say it is listening on port 11434.

4. Download a first model

Before the model download (the Ollama pull command, which writes the weights onto the NAS), pick a small tag (3B–7B Q4) that fits the NAS RAM. Choose a first model explains instruct tags, Q4 quantization, and “out of memory” errors.

From the DSM Terminal (or an admin SSH session):

docker exec -it ollama ollama pull llama3.2:3b

Match the container name from docker ps, and change the tag if you picked another model in the guide above.

5. Test the API from a PC on the LAN

Expected result: a JSON list of models, then a short text reply. The sample prompt asks for a one-sentence hello in French, so the reply should be one short sentence. If a 3B model takes several seconds, write that down for the comparison below.

On a computer on the same network (not from the internet), with curl:

curl http://IP_DU_NAS:11434/api/tags

A JSON reply that lists models means the API is up. Then a generation test. Replace IP_DU_NAS with the NAS address:

curl http://IP_DU_NAS:11434/api/generate -d '{
  "model": "llama3.2:3b",
  "prompt": "Dis bonjour en une phrase.",
  "stream": false
}'

Time the wait before the first token, and the total duration. That is your NAS baseline.

6. Open WebUI beside it (optional)

Many households want a web interface. Two approaches:

  • A second Open WebUI container pointed at http://IP_NAS:11434. See Docker, Ollama, and Open WebUI for the OLLAMA_BASE_URL variable.
  • Open WebUI on a mini PC or a Raspberry Pi, with Ollama only on the NAS as the API backend.

That split keeps the interface snappy even when the NAS is slow to generate.

Volumes, backups, and updates

  • Leave the DSM system volume alone. Put models in a dedicated shared folder.
  • Hyper Backup or snapshots: include ollama-models if you would rather not download 20 GB again after a failure.
  • Updates: stop the container cleanly, pull the new ollama/ollama image, and start it again. Read the Ollama release notes for breaking changes.
  • Antivirus indexing or Cloud Sync on the models folder: turn them off if the scans slow every model download (pull).

Latency: NAS vs mini PC (rough scale)

Rough scale for a household: the table below is not a benchmark of your NAS. The figures are there to compare realistic setups. They are not a manufacturer promise.

Indicative measurements (a 7B Q4 model, a prompt of about 100 tokens, a reply of about 150 tokens, gigabit Ethernet, October 2026). A DS923+ is a recent 4-bay Synology NAS (an embedded AMD CPU, no dedicated GPU). A 7840HS is a common mobile Ryzen in mini PCs. An RTX 4060 is a consumer NVIDIA card with about 8 GB of VRAM. Your results will move with DSM load and with the disks.

Scenario Typical DS923+ NAS (CPU only) Mini PC, Ryzen 7840HS + iGPU Mini PC + RTX 4060
Time to first token, 3B Q4 2–8 s 0.3–1 s 0.2–0.6 s
3B token rate 8–25 tok/s 40–80 tok/s 60–120 tok/s
Time to first token, 7B Q4 5–20 s 0.5–2 s 0.3–1 s
7B token rate 3–12 tok/s 25–50 tok/s 50–90 tok/s
Downloading a ~4 GB model Limited by the NAS hard drive or SSD Fast NVMe Fast NVMe
Noise under load Moderate NAS fans Varies Louder

Rough scale, October 2026. Measure your own NAS with curl in step 5 (the API test).

Quick read (beginner): the NAS is enough for calls that are spaced out (home automation, scripts, short summaries). Move the main inference to a mini PC or a Mac mini as soon as several people chat one after another, or as soon as code completion (Continue) needs to stay smooth.

A closer read: a home-automation question such as “close the blinds if…” every ten minutes is a NAS job. A teenager sending homework messages one after another will be happier on a mini PC or a Mac mini.

A hybrid layout is often the best of both:

  • NAS: files, backups, and a small Ollama for internal APIs;
  • Mini PC: the main inference, Open WebUI, and heavier models.

Also compare noise and power around the clock if you are unsure about a second machine left on all the time.

Safety on a home NAS

  • DSM firewall: allow 11434 only from the LAN subnet, or from the home-automation VLAN.
  • Skip a direct QuickConnect path to Ollama unless you have a reverse proxy plus authentication (Authelia, Synology Application Portal, and similar).
  • Admin accounts: two-factor authentication on DSM. Use a non-admin Docker user for day-to-day upkeep.
  • DSM updates: security patches come first. Ollama is not a reason to leave DSM behind.

Uses that hold up

  1. A Home Assistant webhook: one sentence of weather summary, or a plain-language instruction. Five seconds of latency is acceptable.
  2. n8n: pull JSON fields out of an email. A small model, processed in a batch (several files overnight, with nobody waiting).
  3. A model server for several LAN clients: it centralizes GGUF storage, and each client can also infer locally if you later move the work to a mini PC.

Edge cases: a long Whisper job on the NAS (CPU only), embeddings for a whole document library (numeric fingerprints used for search), and ComfyUI. Prefer a machine with a GPU (VRAM).

Quick troubleshooting

Symptom What to try
connection refused The container is stopped, the port is wrong, or the DSM firewall is blocking it
Very slow from a 3B model RAM is full. Close containers, add RAM, or use a smaller model
Disk full Remove old tags with ollama list and ollama rm
Model download error (pull) Not enough space, or NAS DNS. Try docker exec … ping registry-1.docker.io
OOM kill Raise the container RAM limit, or use a smaller model

Conclusion

Ollama on Synology is a smart add-on to family storage. It is not the core of a fast AI assistant for the whole household. Install it with Container Manager, keep the models on an SSD, test the API with curl (step 5), and measure latency honestly against a mini PC. If the wait is annoying, move the main inference to a quiet mini PC and keep the NAS for a light API and for files. For the wider hardware picture: choose hardware for home AI and the hardware selector (indicative profiles, October 2026 ranges to check against the market).

Shopping after this guide (NAS upgrade or a move to a mini PC)

A sensible order: (1) measure latency on your NAS (step 5), (2) read the latency comparison, (3) only then look at the market. The links open an indicative amazon.co.uk search. An affiliate commission is possible, at no extra cost to you. Price, stock, and DSM compatibility still have to be checked on each product page (how affiliation works, October 2026).

What you need after this guide Indicative search Reminder
SODIMM RAM to upgrade the NAS SODIMM RAM on amazon.co.uk Cross-check the Synology chassis maximum RAM before you order
A mini PC for the main inference mini PCs on amazon.co.uk A hybrid layout: the NAS for storage and a light API, the mini PC for a chat that keeps up

Common questions

Can I run Ollama on a NAS only, with no mini PC?

Yes for small models (3B–7B) and slow internal API calls. No for a family assistant that feels as responsive as a mini PC with a GPU.

Is my DS218+ (2 GB of RAM) enough?

Not for serious Ollama use: too little RAM, and a weak CPU. At best, a one-off experiment with a 1B–3B model after a RAM upgrade, if the chassis allows it.

What is the minimum Synology for Ollama in Docker?

An x86_64 model (recent Plus series: DS920+, DS923+, DS224+), 8 GB of RAM to start, and 16–32 GB if DSM also hosts Synology Photos, Docker, and other services.

Can I use the NAS Intel Quick Sync GPU?

In practice, no, for standard Ollama in Docker on Synology. There is no documented GPU passthrough the way a PC has it. Inference stays on the CPU.

Ollama native, without Docker, on Synology?

There is no official Synology package. The supported path is Container Manager (a container). A community SPK is outside this guide, because you would be on your own for updates.

Should I expose the Ollama port on the internet?

No, for a household. Keep it on the LAN or a Synology VPN. Skip port forwarding for 11434 unless you have a reverse proxy and authentication. Keep guest Wi-Fi away from port 11434.

Are entry-level ARM Synology models a fit?

They are very limited. amd64 Docker images are emulated or missing. Save the idea for a recent Intel or AMD NAS.

Where do I store the GGUF weights?

A dedicated volume on an SSD, or an SSD cache, not on slow hard drives alone. See Volumes, backups, and updates.

Does the NAS replace a ChatGPT subscription for everyday chat?

For automations, partly. For a conversation that feels like a cloud chat, no. See the 3-year cost comparison.

Sources

Narrow it down with the hardware selector · French version of this section