Best AI NAS for running local LLMs

How much memory do you need for a local LLM?

Size the NAS to the models you want to run, with room to spare. At 4-bit quantization, a model needs roughly 0.6GB per billion parameters, and the NAS operating system and storage services need around 8GB on top.

Longer conversations and documents use more memory for context, so leave headroom.

Does a NAS need a GPU to run an LLM?

No, but it makes a large difference. Without a GPU, the processor runs the model, and speed is limited by memory bandwidth. Small models answer at a comfortable pace; 32B class models answer slowly. A discrete graphics card with 16GB or more of its own memory runs those models many times faster. Look for a PCIe x8 or x16 slot, or an OCuLink port for an external dock.

Which software do these systems run?

Ollama is the most common way to run local models and Open WebUI gives it a chat interface. UGREEN, QNAP and ASUSTOR document both, and on Synology they run through Container Manager. UGREEN's iDX models and Zettlab's Ultra models also ship their own built-in assistants.

Common questions

What is the cheapest NAS that can run a useful local LLM?

A system with 32GB of memory and a recent processor will run 8B class models such as Llama 3.1 8B or Qwen3 8B well enough for a private assistant. Check the under £1,000 list for current prices.

Do NAS NPUs make Ollama faster?

Not yet. NAS NPUs accelerate the manufacturer's photo, search and transcription features. Ollama and llama.cpp run on the CPU or a GPU.