AI NAS Compared: Which Storage Box Can Actually Run AI?

QNAP and UGREEN network storage with NPUs or PCIe GPU slots, plus a computed matrix of which LLMs fit a NAS-installed GPU. Every spec comes from the vendor page linked in the table.

What is an AI NAS?

An AI NAS is a network-attached storage box that either ships a built-in NPU for image and video recognition, or provides a PCIe expansion slot that accepts a GPU card. The two kinds of "AI" are different: QNAP's TS-AI642 (6 TOPS NPU) and TVS-AIh1688ATX (36 TOPS NPU) accelerate photo tagging, face recognition, and surveillance analytics on the box itself, while UGREEN's DXP8800 Plus and DXP4800 Plus expose a PCIe x8 slot so you can install a low-power graphics card for real model inference.

Can a NAS run an LLM?

Only a NAS with a GPU installed in its PCIe slot can run LLMs — a NAS NPU cannot. The NPUs in current AI NAS boxes are INT8 vision accelerators rated in TOPS for image models, not FP16 tensor hardware for transformers; a 36 TOPS Intel AI Boost NPU will not run Llama or DeepSeek, it runs QNAP's QuMagie photo recognition and QVR face analytics. For language models you install a low-wattage GPU (≤200 W class) in the slot, give it enough VRAM for the weights, and run llama.cpp or vLLM yourself — the NAS then serves as both storage and inference host.

AI NAS comparison table

Specs from vendor pages (QNAP product pages, UGREEN store). The QNAP TVS-AIh1688ATX's three Gen4 slots and 550 W PSU give it the most expansion headroom; UGREEN lists its x8 slot for GPU and NIC cards. QNAP positions the TVS slots primarily for NIC/NVMe/Thunderbolt 5 cards — check the vendor's compatibility list before committing to a specific GPU.

Which LLMs fit a GPU installed in a NAS?

The matrix below computes whether a model fits one low-power GPU (the realistic NAS configuration) at Q4_K_M quantization with 4k context — required VRAM = weights + 10% loading overhead + KV cache, the same math as our model pages. Card list: current consumer GPUs from our database with VRAM ≥8 GB and TDP ≤200 W. Green = fits on that single card.

Weights-only math at Q4_K_M (4.8 bits/weight) + overhead + 4k KV cache. Real serving also needs activation memory and leaves headroom for context growth, so treat "Fits" as the minimum viable configuration.

What is the realistic LLM ceiling for a NAS GPU?

An 8B-parameter model like Llama 3.1 8B (about 5.8 GB at Q4_K_M) runs on any 8 GB+ card in the table, which is why single-GPU NAS builds standardize on it. A 32B model needs roughly 21 GB and exceeds every 8–16 GB card here — no current low-power card reaches it; the entry point is a used 24 GB card (RTX 3090, 350 W) whose power draw and triple-slot cooler exceed a typical NAS PSU and airflow budget. 70B and DeepSeek V4-scale models are multi-GPU workloads that belong in a workstation or server, not a NAS enclosure. For those, see our server accelerator matrix instead.

NPU NAS vs GPU NAS: which do you need?

Buy an NPU NAS (TS-AI642, TVS-AIh1688ATX) if the AI you want is what the vendor ships: faster photo library tagging, face recognition for surveillance, people counting — services that run inside QNAP's apps with zero setup. Buy a PCIe-slot NAS (DXP8800 Plus, DXP4800 Plus, or the TVS with its three Gen4 slots) if you want to run your own models, because the slot plus a 115–180 W GPU turns the box into a combined storage and inference server. The two UGREEN boxes have no NPU at all: their "AI" features run on the CPU and integrated graphics, so image recognition is slower there than on QNAP's NPU models, while GPU expansion is stronger than the ARM TS-AI642, whose PCIe slot tops out at network and QM2 cards.

Related

• Edge AI devices (Jetson) compared
• AI server GPUs and multi-GPU model-fit matrix
• Best GPUs for LLMs guide
• VRAM calculator (any model, any quant)
• Can-you-run-it model pages