Intermediate
NAS for AI: Training Data & Model Storage
Set up a NAS as central storage for your AI lab. Share model files, training data, and backups across machines over your LAN.
What You'll Need
- A 2-bay (or more) NAS — Synology, TerraMaster, or UGREEN
- 2x hard drives — 4TB+ each recommended (HDD for bulk, SSD for speed)
- An Ethernet cable — NAS over WiFi is painfully slow; wire it to your router
- A computer on the same local network
💡 Why a NAS for AI? Models are big. Llama 3 70B is 40GB. Training datasets can be 100GB+. A NAS gives you one place to store everything, accessible from any machine on your network — your AI server, your laptop, your Pi. No more copying files between USB drives.
1 Choose Your NAS
| NAS | Bays | RAM | Network | Price | Best For |
|---|---|---|---|---|---|
| UGREEN NASync DXP2800 | 2 | 4GB | 2.5GbE | ~$200 | Budget-friendly starter NAS |
| TerraMaster F2-223 | 2 | 4GB | 2.5GbE | ~$230 | Good value, Docker support |
| Synology DS224+ | 2 | 2GB | 1GbE | ~$300 | Best software ecosystem (DSM) |
| Synology DS423+ | 4 | 2GB | 2x 1GbE | ~$500 | Room to grow, expandable |
| QNAP TS-264 | 2 | 8GB | 2.5GbE | ~$400 | Power user, HDMI out, more RAM |
💡 For most people: A 2-bay NAS with 2x 4TB drives in RAID 1 (mirror) gives you 4TB of usable, redundant storage for ~$350 total. That's enough for dozens of large models plus training data.
⚠️ RAID is not backup. RAID 1 protects against a single drive failure, but it won't save you from accidental deletion, ransomware, or the NAS itself dying. Always keep important data in at least two places.
2 Initial Setup & Storage Pool
After inserting your drives and connecting Ethernet:
- Find your NAS on the network — most brands have a discovery tool:
# Synology: find.synology.com # TerraMaster: tnas.online # UGREEN: ugreenlink.com # Or just scan your network: arp -a | grep -i "unknown"
- Run the setup wizard — create your admin account and let it initialize the drives
- Create a storage pool — choose RAID 1 (mirror) for redundancy, or SHR (Synology Hybrid RAID) if available
- Create shared folders — set up folders for different purposes:
# Recommended folder structure: /ai-models/ # Ollama models, GGUF files /training-data/ # JSONL, datasets, corpora /backups/ # Machine backups, configs /projects/ # Shared project files
💡 Enable SMB and NFS. SMB (Samba) works best for Mac and Windows. NFS is faster for Linux. Most NAS devices support both — turn them both on in the settings.
3 Mount the NAS from Your AI Machines
Once your NAS has shared folders, mount them on every machine that needs access.
macOS — mount via Finder or terminal:
# Create a mount point
sudo mkdir -p /Volumes/nas-ai
# Mount the SMB share (replace NAS_IP and folder name)
mount_smbfs //your-user@192.168.1.50/ai-models /Volumes/nas-ai
# Verify
ls /Volumes/nas-ai
Linux — mount via fstab for auto-mount on boot:
# Install CIFS utilities
sudo apt install cifs-utils
# Create mount point
sudo mkdir -p /mnt/nas-ai
# Create a credentials file (keeps password out of fstab)
sudo nano /etc/nas-credentials
# Contents:
# username=your-user
# password=your-password
sudo chmod 600 /etc/nas-credentials
# Add to /etc/fstab for auto-mount
//192.168.1.50/ai-models /mnt/nas-ai cifs credentials=/etc/nas-credentials,uid=1000,gid=1000 0 0
# Mount now without rebooting
sudo mount -a
Windows — map as a network drive:
# In PowerShell or File Explorer → Map Network Drive:
net use Z: \\192.168.1.50\ai-models /persistent:yes
4 Point Ollama & Training Data at the NAS
Now redirect your AI tools to use the NAS for storage. This means any machine on your network pulls models from the same place.
Ollama — change the model storage path:
# macOS / Linux — set the environment variable
export OLLAMA_MODELS=/Volumes/nas-ai/ollama
# Make it permanent (add to your shell profile)
echo 'export OLLAMA_MODELS=/Volumes/nas-ai/ollama' >> ~/.zshrc
# Restart Ollama, then pull a model — it goes to the NAS
ollama pull llama3.2:3b
ls /Volumes/nas-ai/ollama/manifests/
⚠️ Network speed matters. Loading a 7B model from a 1GbE NAS takes ~5 seconds. From a 2.5GbE NAS, ~2 seconds. Once loaded into RAM, inference speed is the same — the NAS only affects initial load time. For very large models (70B+), consider caching them locally and using the NAS as the source of truth.
Symlink approach — keep tools unchanged, redirect the folder:
# Instead of changing env vars, symlink the default location
# macOS default: ~/.ollama/models
# Linux default: /usr/share/ollama/.ollama/models
# Move existing models to NAS first
mv ~/.ollama/models/* /Volumes/nas-ai/ollama/
# Create symlink
ln -s /Volumes/nas-ai/ollama ~/.ollama/models
Training data — organize on the NAS:
# Recommended training data structure
/training-data/
├── finetune/
│ ├── identity_train.jsonl
│ ├── reasoning_train.jsonl
│ └── combined.jsonl
├── datasets/
│ ├── wikipedia-en/
│ └── code-alpaca/
└── exports/
├── chat_history.jsonl
└── eval_results.json
💡 This pairs with the Model Library guide. That guide covers organizing models on a local NVMe. This guide extends that to network storage — useful when you have multiple machines that all need the same models.
5 Backups & Maintenance
A NAS isn't just storage — it's your safety net. Set up automated backups and health monitoring.
Automated snapshots (Synology):
# Synology → Snapshot Replication → Scheduled Task
# Set daily snapshots on /ai-models/ and /training-data/
# Retention: keep last 7 daily, last 4 weekly
# This lets you roll back if a model gets corrupted or deleted
Scrub schedule — check drive health:
# Most NAS UIs have a "Data Scrubbing" option under Storage
# Schedule monthly — it reads every block and verifies checksums
# Catches silent data corruption (bit rot) before it becomes a problem
Backup the NAS itself:
- External USB drive — plug into the NAS USB port, schedule nightly copy of critical folders
- Second NAS — rsync between two NAS devices for full redundancy
- Cloud (optional) — Synology Hyper Backup supports Backblaze B2 (~$5/TB/month) for offsite
Monitor drive health from the command line:
# SSH into your NAS (enable SSH in settings first)
ssh admin@192.168.1.50
# Check SMART status on Synology
smartctl -a /dev/sda
# Key fields to watch:
# Reallocated_Sector_Ct — should be 0
# Current_Pending_Sector — should be 0
# Temperature_Celsius — under 45°C
| Storage Task | Size Estimate | 2TB NAS | 4TB NAS | 8TB NAS |
|---|---|---|---|---|
| Small models (1B–3B) | ~2GB each | ~1000 | ~2000 | ~4000 |
| Medium models (7B–8B) | ~5GB each | ~400 | ~800 | ~1600 |
| Large models (70B) | ~40GB each | ~50 | ~100 | ~200 |
| Fine-tune JSONL datasets | ~10MB–1GB | Thousands | Thousands | Thousands |
| Chat history exports | ~1MB/1000 msgs | Millions | Millions | Millions |
✅ What You've Set Up
- A NAS with a RAID-protected storage pool and organized shared folders
- Network mounts on Mac, Linux, or Windows for seamless access
- Ollama models stored centrally — pull once, use from any machine
- Training data and exports organized and accessible over LAN
- Automated snapshots, scrub schedules, and drive health monitoring
Next Steps
- Run Docker on the NAS — some NAS devices (Synology, QNAP) can run containers directly. Host Open WebUI right on the NAS.
- Set up a Time Machine target — most NAS devices can act as a Mac backup destination. Protect your whole machine, not just AI files.
- Add an SSD cache — if your NAS has M.2 slots, adding an SSD cache dramatically speeds up small file reads (metadata, JSONL files).
- Pair with your AI server — the always-on AI server loads models from the NAS, serves them to everything else on your network.
- Centralize training data for fine-tuning — keep all your JSONL files on the NAS and point your model library at it.