How is the memory figure calculated?
Three parts add up to the minimum. Model weights are the published file size of the model at your chosen quantization, or, for a custom size, the parameter count multiplied by the bits each weight uses. The context cache stores the conversation so far: it grows with the context length and with every person chatting at the same time. Runtime overhead adds about 2GB for the software running the model.
The comfortable figure adds 20% headroom and 8GB for the NAS operating system, storage services and other apps. The recommended figure rounds that up to the next common memory size, such as 32GB, 64GB or 96GB.
How is the context cache estimated?
Each word in the conversation is stored as a small block of numbers for every layer of the model. Most current models use a design called grouped-query attention, which keeps that cost down. We use typical figures by model size: about 0.06MB per token for models up to 4B, 0.13MB up to 9B, 0.2MB up to 15B, 0.25MB up to 35B, 0.32MB up to 80B and 0.5MB above that. A 32B model with an 8K context uses roughly 2GB of cache for one person (Estimated).
What context length should I choose?
For short chats, 4K is plenty. For pasting documents or long conversations, 16K to 32K is more realistic. Very long contexts such as 128K use a lot of memory, especially for larger models, and only make sense when you need to process long documents in one go.
How many users should I enter?
The number of people likely to use the model at the same moment, not the number of accounts. A household rarely has more than one or two at once; a small office might have three to five.
Which quantization should I pick?
Q4 is the usual choice: about a quarter of the original size with little loss in quality. Pick Q5 or Q8 if you have memory to spare and want the best answers; Q3 or Q2 only when nothing larger fits.
Worked examples
Llama 3.1 8B at Q4, 8K context, one user: about 4.9GB of weights, 1GB of context cache and 2GB of runtime, a minimum of around 8GB. With headroom for the NAS, a 24GB system is comfortable (Estimated).
Qwen3 32B at Q4, 8K context, one user: about 20GB of weights, 2GB of cache and 2GB of runtime, a minimum of around 24GB. The comfortable figure is about 37GB, so a 48GB or 64GB system is the right size.
Llama 3.3 70B at Q4, 8K context, two users: about 43GB of weights and 5GB of cache, a minimum near 50GB. The comfortable figure is about 68GB, which calls for a 96GB-class NAS, and patience without a GPU. Dropping to one user or a 4K context brings the cache down sharply, but the weights alone still rule out a 32GB or 48GB system.
How do I use the result?
Buy a NAS whose maximum supported memory is at least the recommended figure, and check whether that memory can be upgraded later. The link under the result opens the Finder with that memory as a minimum, and the table lists every system that supports enough memory, cheapest verified UK price first.
Common questions
Does the calculator account for GPU memory?
No. It sizes system memory. With a graphics card, the model should fit in the card's own memory for full speed.
Why is my result higher than the model's file size?
Because the model needs room for the conversation, the software running it and the NAS itself, not just its weights.