How does the check work?
Every model in our list has a published file size at each quantization, taken from the Ollama library. Every NAS has a maximum supported memory from its manufacturer, or a fixed amount when memory is soldered. The check compares the two.
A model runs comfortably when the NAS has at least the model's size multiplied by 1.2 plus 8GB. The extra 20% covers the conversation context and the runtime; the 8GB covers the NAS operating system, storage services and other apps. A model runs with limitations when it only fits with little headroom, when only a heavily compressed version fits, or when it is large enough to be slow on a NAS without a GPU path. It does not fit when even the smallest published version needs more memory than the NAS supports.
What is quantization and why does it matter?
Quantization stores a model's weights with fewer bits. Q8 keeps close to full quality at about half the size of the original; Q4 is the usual balance, roughly a quarter of the size, with a small loss in quality; Q3 and Q2 shrink further and lose more. The check picks the highest quality that fits comfortably, so you see the best version your NAS can run.
Why is there no speed figure?
Speed depends on memory bandwidth, the processor, the graphics, the runtime and the length of the conversation, and published claims vary widely. We show a tokens-per-second figure only when it has been measured on that system, by AINAS or a credible source. Until then the result names the likely bottleneck instead.
What does the GPU situation mean?
A graphics card holding the whole model runs it many times faster than a processor. If the NAS has a PCIe slot or OCuLink port, a card can be added later; if it does not, the processor and integrated graphics do all the work. NAS NPUs are not used by common runtimes such as Ollama, so they do not change the answer.
How accurate is it?
The memory calculation is reliable for deciding whether a model will load. Real-world behaviour can differ: very long conversations use more memory, several people chatting at once multiply the context, and other apps on the NAS compete for the same memory. If a result is marginal, plan for more memory or a smaller model.
Which NAS details matter most?
Maximum supported memory matters most, because it decides which models load at all. Whether memory is soldered matters next: a soldered system keeps the amount it ships with, so buy enough at the start. After that, a PCIe slot or OCuLink port for a graphics card decides how fast the larger models will run, and the processor's core count decides how comfortable a CPU-only setup will feel. Drive bays and network speed matter for storage, but have little effect on running a model.
What if my NAS cannot run the model?
The result suggests an alternative system that runs it comfortably, and lists every system that does with its current UK price. You can also try a smaller model in the same family, which often gives most of the quality at a fraction of the memory.
Common questions
Does the check include GPU memory?
No. It assumes the model runs from system memory. With a graphics card fitted, the card's memory decides which models run at full speed.
Why does a soldered-memory NAS use its fitted amount?
Because it cannot be upgraded. A 64GB system with soldered memory will always have 64GB.