TrueNAS Scale homelab & hardware
Hardware platform, ZFS storage, self-hosted apps, security tooling and AI-related services.
Guide and documentation by Alban Andrieu.
Hardware, apps, ZFS, automation and AI — open-source and hands-on.
TrueNAS services
Self-hosted applications on TrueNAS Scale.
Service icons are vendored from selfh.st/icons. For more open-source self-hosted ideas, see selfh.st/apps.
Homelab for security, AI and freelance projects
This homelab was built to host upcoming Nabla projects and services.
Its purpose is to test security and AI service integrations and reproduce attacks targeting my domains.
After seven years of reliable service, the previous FreeNAS system failed: its motherboard, CPU and several hard drives were faulty. It was time for an upgrade.
One physical Mini-ITX system running TrueNAS Scale: dependable ZFS storage and a homelab host for applications, home automation, security tooling and compute-intensive workloads, including AI experiments.
- ZFS and datasetsPools, snapshots and replication
- Homelab hostApplications, VMs and internal services
- Security stackSecurity tooling and monitoring
- AI and computeAM5 capacity for models and jobs
Bill of materials
Reused components
- Case: previous iXsystems chassis, approximately 170 × 170 mm, Mini-ITX format.
- Power supply: previous 200 W unit, incompatible with the MSI motherboard’s 8-pin CPU connector.
- Kingston SKC600MS mSATA SSD — 256 GB TLC 3D NAND. €83.
- Storage: 4 × Western Digital Red WD30EFRX — 3 TB each, SATA 6 Gb/s.
New purchases
- MSI MPG B650I Edge WiFi — Mini-ITX AM5 motherboard, DDR5, PCIe 4.0, M.2 and Wi-Fi 6E. €174.67.
- ORICO mSATA SSD enclosure (USB 3.0, 5 Gbps) — Tool-free USB 3.0 enclosure for 50 × 30 mm mSATA SSDs, up to 2 TB. €17.99.
- Noctua NH-L9a-AM5 — Low-profile CPU cooler for AMD AM5. €49.90.
- AMD Ryzen 7 7700 — 3.8 GHz processor with 32 MB L3 cache. €212.75.
- Crucial DDR5 64 GB 5600 MHz — CL46 memory (CT64G56C46U5; also supports 5200/4800 MHz). €679.00.
- ATX 500 W PSU — ATX 500 W power supply; the smaller ordered unit was never delivered. €50.00.
Total price€1,184.31Sum of motherboard, enclosure, CPU, RAM, CPU cooler and power supply.
Possible future upgrades
Fast SSD pool for services
Planned2 × WD Red SN700 1 TB NVMe SSD — €299.99 each, €599.98 total.
boot-poolMirror → fastpoolRAIDZ → cpoolExisting application datasets will be migrated selectively to the SSD pool. For example, /mnt/cpool/openwebui can become /mnt/fastpool/appdata/openwebui, while bulk data remains on cpool.
GPU options for local AI inference
The motherboard provides one PCIe x16 slot. The final choice is primarily constrained by chassis clearance, PSU capacity and the VRAM required by the models.
Price snapshot: 24 August 2026; retail prices can change quickly.
PNY NVIDIA RTX PRO 4000 Blackwell SFF Edition
Preferred option- VRAM
- 24 GB GDDR7 ECC
- Power
- 70 W
- Form factor
- 168 mm, dual-slot, low-profile
- Observed price
- €2,483.00
Best fit for the existing compact chassis and 500 W PSU. Its 70 W board power, low-profile dual-slot format and 24 GB ECC VRAM make it the preferred inference upgrade.
NVIDIA RTX PRO 5000 Blackwell
- VRAM
- 48 GB GDDR7 ECC
- Power
- 300 W
- Form factor
- 267 mm, dual-slot, full-height
- From
- €7,172.61
Doubles VRAM to 48 GB and greatly expands fully-GPU model capacity, but its 300 W power draw and full-height 267 mm format likely require a larger chassis and stronger PSU.
NVIDIA RTX PRO 6000 Blackwell Max-Q
- VRAM
- 96 GB GDDR7 ECC
- Power
- 300 W
- Form factor
- 267 mm, dual-slot, full-height
- From
- €18,888.97
96 GB ECC VRAM is the maximum local-inference target in this roadmap. Assume a new chassis, stronger PSU and carefully designed airflow for the 300 W Max-Q card.
Local model capacity with RTX PRO 4000 SFF
Indicative capacity for a 24 GB GPU. Measured speeds use public benchmarks on this exact GPU where available; other figures are conservative engineering estimates and depend on quantization, context length and inference engine.
Models that fit in 24 GB VRAM
| Model | Approx. model memory | Best use | Generation speed | Practical parallel prompts |
|---|---|---|---|---|
| gpt-oss-20b MXFP4 | 11.27 GiB | General reasoning, agents, tool use and coding | 109.7 tok/smeasured on RTX PRO 4000 Blackwell SFF | 4–8 |
| Qwen3.6-35B-A3B-MTP IQ4_NL | 17.25 GiB | Agentic coding, repository work, general multimodal reasoning | 95.9 tok/smeasured on RTX PRO 4000 Blackwell SFF | 2–4 |
| Qwen3-Coder-30B-A3B-Instruct Q4_K_M | ~18.6 GB | Dedicated coding and repository-scale development | ~70–100 tok/sestimated | 2–4 |
| MADLAD-400-10B-MT 4-bit | ~6–8 GB | High-volume multilingual translation (400+ languages) | ~30–60 tok/sestimated | 8+ |
Larger hybrid models with 128 GB system RAM
| Model | Approx. model memory | Best use | Generation speed | Practical parallel prompts |
|---|---|---|---|---|
| gpt-oss-120b MXFP4 | 63.4 GB | Higher-quality general reasoning, agents and coding | ~20–30 tok/sestimated | 1–2 |
| Llama 3.3 70B Instruct Q4_K_M | ~42 GB | Large multilingual general-purpose model; dense CPU/GPU offload | ~2–4 tok/sestimated | 1 |
| NVIDIA Nemotron 3 Super 120B-A12B (quantized) | ~65 GB | Agentic reasoning, RAG and long-context experimentation; quantized community runtimes | ~8–16 tok/sestimated | 1 |
Parallel figures assume one shared llama.cpp/vLLM-style server with continuous batching and short-to-moderate contexts. They do not mean loading a separate copy of the model for every user.
System RAM does not become VRAM. Larger models use CPU/RAM offload and are therefore substantially slower than models that stay entirely on the GPU. On TrueNAS, the same 128 GB is also shared with ZFS ARC and other applications, so very large models need an explicit memory budget and sensible ARC/container limits.
