AI NAS, exposed: built-in, nominal, or marketing?
"AI NAS" hides three different machines. We grade each current model on what it actually ships — memory, memory ceiling, iGPU/NPU — against what LLMs need, so you can tell the genuinely-AI ones from the name-only ones. Facts as of 2026-11.
First, the three different "AI NAS" meanings
A box tuned for ② (an NPU, 8 GB soldered) can be the worst choice for ③. That mismatch is the whole story below.
The verdict table — which are real, which are name-only
| Model | CPU | Shipped RAM | RAM ceiling | GPU / NPU | Verdict |
|---|---|---|---|---|---|
| DS925+ | AMD Ryzen V1500B (4C/8T) x86 | 4 GB ECC (2 slots, expandable to 32 GB) | 32 GB | no discrete GPU, no iGPU in spec, no NPU | 名义 AI 名义 AI，要加内存 |
| DS1825+ | AMD Ryzen V1500B (4C/8T) x86 | 8 GB ECC (4 slots, expandable) | 见官方（多槽可插） | no NPU | 名义 AI 名义 AI，要加内存 |
| TS-464 | Intel Celeron N5095 (4C/4T) x86 | 8 GB SODIMM DDR4 (2 slots) | 16 GB (2 x 8GB) | Intel UHD iGPU（可借给推理） <span style="color:#d9534f">NPU</span> | 轻度原生 AI 轻度原生 AI |
| TS-473A | AMD Ryzen V1500B (4C/8T) x86 | 8 GB SODIMM DDR4 | 64 GB (2 x 32 GB) | 无 NPU | 原生 AI 主力 原生 AI 主力 |
| TS-673A | AMD Ryzen V1500B (4C/8T) x86 | 8 GB SODIMM DDR4 | 64 GB (2 x 32 GB) | 无 NPU | 原生 AI 主力 原生 AI 主力 |
| TVS-AIh1688ATX | Intel Core Ultra 9 24 核 (8P+16E) + 24 线程 x86 | 48 GB ECC DDR5 | 192 GB (4 x 48 GB) | Intel Arc iGPU（可借给推理）+ 支持加独显 <span style="color:#d9534f">NPU</span> | 天生 AI 旗舰 天生 AI 旗舰 |
| TS-AI642 | 8 核 ARM (4x A76 + 4x A55) ARM | 8 GB 板载（不可扩） | 8 GB | Mali-G610 iGPU + 6 TOPS NPU <span style="color:#d9534f">NPU</span> | AI 陷阱款 AI 陷阱款：NPU 不跑 LLM |
| DXP6800 Pro | Intel Core i5-1235U x86 | 见官方规格页（DDR5，可扩展） | 见官方 | Intel Iris Xe iGPU（NAS 里最强集成显卡） <span style="color:#d9534f">NPU</span> | 原生 AI 主力 原生 AI 主力（iGPU 强） |
| DXP4800 | Intel N100 (4C) x86 | DDR5（官网确认，量见规格页） | 见官方 | Intel UHD iGPU <span style="color:#d9534f">NPU</span> | 轻度原生 AI 轻度原生 AI |
| HA100 (HomeAgent) | 官网未公开规格（AI-first 定位） 见规格页 | 官网无规格页，仅「AI-first design」 | 待官网 | 官网未列 <span style="color:#d9534f">NPU</span> | 营销 AI 款 营销 AI 款：规格未公开 |
| fnOS / FygoOS (DIY x86) | 你装什么板就什么 CPU x86（常见） | 取决于你主板/内存（32-64GB 很便宜） | 取决于板子 | 可上独显/iGPU/NPU 全可选 <span style="color:#d9534f">NPU</span> | 最灵活(DIY) 最灵活：DIY 决定一切 |
The three traps, named
挂着 6 TOPS NPU 很唬人，但 NPU 是给 Frigate/QuAI 视觉用的；8GB 板载不可扩，跑 8B LLM 都紧，14B 无望。买它当 LLM 机=买错。
UGREEN 把 HA100 当 AI 主打卖，但官网至今没给 CPU/内存/NPU 规格页——「能不能跑 8B」无法从官方口径确认。这是「AI NAS 营销 vs 现实」的典型样本。
When you must add hardware
| You want… | Minimum RAM | On a closed NAS | Otherwise |
|---|---|---|---|
| 3B chat | 4-8 GB | Any 8 GB SKU | — |
| 7-8B chat | 16 GB | QNAP/UGREEN i5 (RAM-upgrade) | add an M.2/DDR module |
| 14B | 32 GB | QNAP 64 GB SKUs, UGREEN DDR5 | DIY fnOS box |
| 32-70B | 64-128 GB | TVS-AIh1688 (192 GB) only | DIY + dGPU, or a mini-PC beside the NAS |
| Fast vision | — | iGPU / NPU for light use | discrete GPU or a Frigate box |
AI fit checker
Pick the model you want, then your NAS — the checker compares the model's RAM need against the box's shipped and maximum memory.
FAQ
What does "AI NAS" actually mean?
Marketing usually means one of three different things, and they are not the same: (1) AI photo tagging — a CPU or NPU recognizing faces/places in your photos; (2) NVR AI — an NPU doing motion/object detection for cameras; (3) local LLM — running a chat model. A box tuned for (1) or (2) does not automatically do (3). Always ask which kind.
Is my NAS "AI" if it has no GPU or NPU?
Yes — for a local LLM, RAM matters far more than a GPU or NPU. A plain x86 NAS with 32 GB RAM runs an 8-14B model fine on CPU. A box with an NPU but only 8 GB soldered memory (the QNAP TS-AI642 pattern) may be worse for LLMs than a boring NAS you upgraded to 32 GB. NPU accelerates the wrong workload for LLMs today.
When do I need extra hardware beyond the NAS?
Three tells. If your NAS caps at 8 GB and you want 8B+ models, add RAM or an M.2 module. If you want 32-70B, you need 32-128 GB — usually a DIY build, a high-memory QNAP, or a dedicated mini-PC with a GPU next to the NAS. If you want fast vision or heavy inference, add a discrete GPU or a small GPU box. The NAS holds the data; the compute can live elsewhere.
How do I check if a specific NAS fits a model?
Match the model download size plus headroom to the NAS available RAM. Find the shipped and maximum RAM, subtract what the OS uses, and compare to the model. Our checker below does this for common models and NAS SKUs. When in doubt, the maximum (not shipped) RAM is what a model needs to fit.