Which NAS can run a local LLM?
The answer is a function of model size × RAM × acceleration, not brand. Below: the model-size ladder (quantized sizes per Ollama's public model cards), then current NAS models with a per-model verdict. Facts as of 2026-11; specs from vendor pages.
Step 1 — the model-size ladder
| Size | Common models | Q4 weight | Total RAM | What it feels like on NAS hardware |
|---|---|---|---|---|
| 3B | Llama 3.2 3B, Qwen2.5 3B, Phi-3 mini | ~2 GB | 4-6 GB | usable on CPU alone; typically 8-15 tokens/s on a modern NAS CPU |
| 7-8B | Llama 3.1 8B, Mistral 7B, Qwen2.5 7B | ~4.7 GB | 8-10 GB | CPU-only is slow (3-8 tokens/s); an iGPU/NPU or 16 GB+ RAM changes the experience |
| 13-14B | Qwen2.5 14B, Mistral Nemo 12B | ~9 GB | 12-16 GB | CPU-only works but responses crawl; expect batch/summary workloads only |
| 32B | Qwen2.5 32B, DeepSeek-R1-Distill 32B | ~20 GB | 24-28 GB | realistic only on DIY x86 (fnOS/TrueNAS) with 32-64 GB RAM |
| 70B | Llama 3.3 70B | ~40 GB | 48 GB+ | out of scope for any consumer/half-pro NAS on the market |
Numbers are for 4-bit quantization (the default Ollama download). KV-cache grows with context length, so long-document workloads push the "total RAM" column up.
Step 2 — current NAS models, ranked for local LLMs
| Model | CPU | RAM (as shipped) | Acceleration | 3B | 7-8B | Note |
|---|---|---|---|---|---|---|
| DS925+ | AMD Ryzen V1500B (4C/8T) | 4 GB ECC | none (no iGPU, no NPU) | ✔ yes | ✔ with RAM upgrade | CPU-only inference; no integrated GPU means every token comes from CPU cores |
| DS1825+ | AMD Ryzen V1500B (4C/8T) | 8 GB ECC | none (no iGPU, no NPU) | ✔ yes | ⚠ tight on stock RAM | 8 GB ships close to the 8B ceiling once DSM and services take their share |
| TS-464 | Intel Celeron N5095 | 8 GB (max 16 GB) | Intel UHD Graphics (iGPU) | ✔ yes | ✔ at 16 GB | the cheapest common on-ramp: 16 GB upgrade + iGPU offload |
| TS-AI642 | 8-core ARM (4x A76 + 4x A55) | 8 GB (non-expandable) | 6 TOPS NPU + Mali-G610 | ✔ yes | ⚠ tight on stock RAM | NPU is built for video/image recognition (frigate-class), not LLM tokens; 8 GB is soldered |
| TVS-AIh1688ATX | see QNAP spec page | see QNAP spec page | NPU-equipped AI flagship (16-bay) | see spec | see spec | QNAP's 2026 AI-flagship line; our drive-compatibility page tracks 36 certified drives for it |
| DXP4800 | Intel N100 | 8 GB DDR5 | Intel UHD (iGPU) | ✔ yes | ⚠ tight on stock RAM | low-power entry; UGREEN markets it with AI photo features on UGOS Pro |
| DXP6800 Pro | Intel Core i5-1235U | 8 GB DDR5 | Intel Iris Xe (iGPU) | ✔ yes | ✔ at 16 GB | Iris Xe iGPU is the strongest integrated graphics in the mainstream NAS tier |
| HA100 (HomeAgent) | AI-agent appliance (spec sheet at UGREEN) | spec sheet at UGREEN | AI-first design | n/a | n/a | UGREEN's dedicated AI NAS line; our database already tracks 31 compatible drive entries for it |
| fnOS / FygoOS on DIY x86 | any x86 you choose | whatever you install (32-64 GB is cheap) | depends on board (iGPU/dGPU/NPU all possible) | ✔ yes | ✔ yes | the only route where a 32B model is realistic: no vendor RAM ceiling, GPU slot is yours |
"3B / 7-8B" = comfortable fit for a Q4 model of that class while the box keeps doing NAS duty. Verdicts assume Docker + Ollama on x86; the hardware-requirements page explains the physics behind each verdict.
Reading the table honestly
The RAM ceiling is the whole game on closed NAS. Synology's 2025 generation ships 4-8 GB ECC; QNAP's TS-AI642 solders 8 GB; UGREEN's DXP Pro models start at 8 GB DDR5 but the platform takes more. Once you subtract DSM/QTS/UGOS and file services, a stock 8 GB box has room for a 3B model and not much else.
iGPU beats NPU for LLMs (today). The NPUs shipping in NAS hardware — QNAP's 6 TOPS in the TS-AI642 — are vision accelerators. For token generation, Intel's Iris Xe (UGREEN i5 models) or any dGPU in a DIY build does the work. See the QNAP AI page for what the TS-AI NPU is actually for.
The DIY route wins above 14B. No closed consumer NAS SKU ships the 32 GB+ that 32B models want. A fnOS box on standard x86 hardware with 32-64 GB is the cheapest realistic path — and it can still run the same Docker stack. Our fnOS drive guide covers the storage side of such a build.
FAQ
Can a NAS run a local LLM?
Yes — any x86 NAS with enough RAM can run Ollama-class models in Docker. The real question is which model size: 3B models fit almost any 8 GB box, 7-8B wants 8-16 GB, 14B wants 16 GB+, and 32B is realistic only on DIY builds (fnOS/TrueNAS) with 32 GB or more. Closed NAS SKUs mostly ship 4-8 GB and cap there.
Does a NAS NPU help run LLMs?
Not the NPUs currently in NAS hardware. The QNAP TS-AI642’s 6 TOPS NPU accelerates image/video recognition — the workload it was designed for. LLM token generation still runs on CPU/GPU paths, and tooling for NPU-offloaded LLMs on NAS SoCs is not vendor-supported.
How much RAM do I need for a 7B model on a NAS?
Plan for 8-10 GB total for a Q4-quantized 7-8B model including KV-cache — which means a stock 8 GB NAS is already tight once the OS and file services take their share. 16 GB makes it comfortable; that usually means a QNAP TS-464 or UGREEN i5 model with a RAM upgrade, or a DIY fnOS box.
Is CPU-only LLM inference usable on a NAS?
For 3B models, yes — modern NAS CPUs manage usable speeds. For 7-8B, CPU-only works but is slow (a few tokens per second); fine for summarizing or batch tasks, frustrating for chat. An iGPU (Intel Iris Xe in UGREEN i5 models) or a discrete GPU in a DIY box changes the experience.