Some links are affiliate links: if you buy through them we may earn a commission, at no extra cost to you. It never changes our assessments. Affiliate disclosure
Most NAS boxes are built for storage: a low-power CPU, a few gigabytes of memory and no capable GPU. They can run tiny models and little else. Two recent exceptions put real AI hardware inside a NAS, and they are very different.
Minisforum N5 MAX: a NAS with Strix Halo
The N5 MAX uses the Ryzen AI Max+ 395, the same chip as the Framework Desktop or the GMKtec EVO-X2 in our guide: 16 Zen 5 cores, a Radeon 8060S GPU and 64 or 128 GB of LPDDR5X shared between CPU and GPU. Minisforum specifies the memory at 8533 MT/s on a 256-bit bus, which is 273 GB/s: about 7% faster than the 256 GB/s of the other Strix Halo machines.
Around that chip it is a proper NAS: five 3.5"/2.5" SATA bays and five M.2 slots (up to 200 TB according to Minisforum), two 10 GbE ports, USB4 v2 and a 250 W built-in power supply. It ships diskless.
What it runs. The same models as any 128 GB Strix Halo box: gpt-oss-120b, Qwen3.5 122B A10B and every mid-size model. With 64 GB you are limited to models up to about 45 GB: Qwen3.6 35B A3B, Gemma 4 31B and smaller. See AMD Ryzen AI Max+ 395 (Strix Halo) for the full table.
Prices. At launch Minisforum asked $2,899 for 64 GB and $3,599 for 128 GB (reported by Tom's Hardware and VideoCardz); the 64 GB model was listed at $2,399 on Minisforum's store when we checked on 12 October 2026. The 64 GB model is on Amazon; the 128 GB one is sold by Minisforum.
Should you buy it? It makes sense if you were going to buy both a NAS and a 128 GB machine: it is one box, one power supply and one network connection. It makes less sense if you want to replace the AI part later: the memory is soldered, and a NAS has to stay on and healthy for years, while AI hardware changes every year. A separate mini PC lets you swap it without touching your data.
Minisforum N5 Pro: a NAS with Ryzen AI 9
The N5 Pro has a Ryzen AI 9 HX PRO 370, a mid-range chip with a 16-CU Radeon 890M and up to 96 GB of DDR5 ECC memory in SO-DIMM slots, plus an OCuLink port that can connect an external graphics card (LinuxLinks review). Its memory is ordinary DDR5 on a 128-bit bus: at most 89.6 GB/s (DDR5-5600, the fastest the chip supports), a third of Strix Halo.
In practice, a user reported 15.3 tok/s with Qwen3-Coder-Next 80B at Q4_K_M and 10.9 tok/s with Qwen3-Next 80B at Q8_0 on a Ryzen AI 9 HX 370 with 96 GB of DDR5 and llama.cpp with Vulkan (report). Those are mixture-of-experts models that read only a few billion parameters per token, which is why they run at a usable pace. Dense models above about 14B will feel slow. See AMD Ryzen AI 9 HX 370 (Strix Point).
Should you buy it? As a NAS that can also host a small assistant (8–30B models, light MoE models), yes. If your goal is local AI first, choose Strix Halo.
The NPU does not matter (yet)
Both chips include a 50 TOPS NPU. llama.cpp and Ollama run on the GPU, not the NPU, so the NPU does not change the numbers above.
Other NAS
We have not found other NAS with unified-memory chips. The usual NAS processors cannot run anything above very small models at a useful speed, so for those, keep the NAS for storage and put the AI on a separate machine.