Run Ollama on a NAS: Check the Memory First
16 of 32 NAS models ship with more memory than the 4.9GB Llama 3.1 8B download
As an Amazon Associate we earn from qualifying purchases. Prices and availability come from Amazon.com as of the last update and may change.
Ollama runs on a NAS as a container, started with the command its Docker page lists under "CPU only", and what limits it there is memory. Ollama's FAQ says a model shown as 100% CPU was loaded entirely in system memory, and that required memory scales with the context window, 4,096 tokens by default. So a model's download size is the first figure to set against the memory in the NAS. Llama 3.2 3B is a 2.0GB download and Llama 3.1 8B is 4.9GB: of the 32 NAS models on record, 16 ship with more memory than the second, and 14 more pass it at the maker's memory limit.
What Ollama's documentation states
| Point | Ollama's wording |
|---|---|
| Running with no graphics card | Under "CPU only": docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama |
| Where the model sits | 100% CPU means the model was loaded entirely in system memory |
| How to check | The ollama ps command shows what is loaded, and its Processor column shows which memory the model was loaded into |
| Context window | By default, Ollama uses a context window size of 4096 tokens; OLLAMA_CONTEXT_LENGTH overrides it |
| Memory and context | Required RAM will scale by OLLAMA_NUM_PARALLEL * OLLAMA_CONTEXT_LENGTH |
| How long a model stays loaded | By default models are kept in memory for 5 minutes before being unloaded |
| Not enough memory for another model | All new requests will be queued until the new model can be loaded |
| Where models are stored on Linux | /usr/share/ollama/.ollama/models, or the directory set in OLLAMA_MODELS |
Download size against NAS memory
| Model | Ollama tag | Download | Context window | NAS models shipping with more memory | More at the maker's limit |
|---|---|---|---|---|---|
| Llama 3.2 1B | llama3.2:1b | 1.3GB | 128K | 32 of 32 | 0 |
| Llama 3.2 3B | llama3.2:3b | 2.0GB | 128K | 26 of 32 | 5 |
| Gemma 3 4B | gemma3:4b | 3.4GB | 128K | 26 of 32 | 5 |
| Mistral 7B | mistral:7b | 4.4GB | 32K | 16 of 32 | 14 |
| Llama 3.1 8B | llama3.1:8b | 4.9GB | 128K | 16 of 32 | 14 |
| Qwen3 8B | qwen3:8b | 5.2GB | 40K | 16 of 32 | 14 |
| Gemma 3 12B | gemma3:12b | 8.2GB | 128K | 1 of 32 | 24 |
| Phi-4 14B | phi4:14b | 9.1GB | 16K | 1 of 32 | 24 |
| Qwen3 14B | qwen3:14b | 9.3GB | 40K | 1 of 32 | 24 |
| Gemma 3 27B | gemma3:27b | 17GB | 128K | 0 of 32 | 15 |
| DeepSeek-R1 32B | deepseek-r1:32b | 20GB | 128K | 0 of 32 | 15 |
| Llama 3.1 70B | llama3.1:70b | 43GB | 128K | 0 of 32 | 7 |
How to read that table
It compares two published numbers, the download size on Ollama's library page and the memory each NAS maker states. It is not a test, and a NAS with exactly as much memory as the download has nothing left for its own system, which is why the count is of models with more memory, not the same.
The Ollama pages read here give no memory requirement per model and no speed for any hardware, so neither appears on this page. The context window column is the model's maximum; Ollama's default is 4,096 tokens whatever the model allows.
Llama 3.1 70B is a 43GB download. No NAS on record ships with that much memory, and 7 can be upgraded past it.
Graphics hardware Ollama lists
| Hardware | What Ollama states |
|---|---|
| Nvidia | GPUs with compute capability 5.0 or higher and driver version 550 or newer |
| AMD Radeon | Supported through the ROCm library; Linux needs the AMD ROCm v7 driver |
| Apple | GPU acceleration through the Metal API |
| Vulkan | Additional GPU support on Windows and Linux, enabled by default when the backend is installed; most Linux distributions need extra components |
What that means for the graphics in a NAS
Ollama's hardware page names Nvidia, AMD Radeon and Apple graphics, and a Vulkan route that links to Intel's and AMD's Linux driver instructions. It does not state that the integrated graphics in any particular processor is supported, so nothing is claimed here for the graphics in any NAS.
Ollama's Docker commands for a graphics card add a --gpus or --device option. The "CPU only" command has neither, and it is the one that needs nothing from the NAS beyond container support.
Related pages
Sizes at every quantization, a calculator for your own unit, and which NAS models state container support.
LLM model size chart · Local AI RAM calculator · Home Assistant on a NAS (container and VM support by model)
Where this comes from
Ollama's FAQ, hardware and Docker documentation and its library tag pages, read on October 12, 2026. NAS memory figures are each maker's, as recorded on this site's device pages.
Checked on October 12, 2026. Sources: Ollama documentation: FAQ; Ollama documentation: Hardware support; Ollama documentation: Docker; Ollama library: llama3.2 tags; Ollama library: llama3.1 tags; Ollama library: gemma3 tags; Ollama library: qwen3 tags; Ollama library: deepseek-r1 tags; Ollama library: mistral tags; Ollama library: phi4 tags.
NAS on Amazon.com, by price
Every new NAS listing tracked, cheapest first. Memory figures on this page are for the models named in its tables only.
Frequently asked questions
Can Ollama run on a NAS without a graphics card?
Yes. Ollama's Docker page lists a CPU-only command, and its FAQ says a model shown as 100% CPU is loaded entirely in system memory.
How much memory does a NAS need for Ollama?
More than the model's download, at least. Llama 3.1 8B is 4.9GB and Llama 3.2 3B is 2.0GB on Ollama's library pages, and Ollama says required memory also scales with the context window.
What is Ollama's default context window?
4,096 tokens, by Ollama's FAQ. The OLLAMA_CONTEXT_LENGTH variable overrides it, and required memory scales with it.
Where does Ollama keep its models in Docker?
In the volume its Docker command mounts at /root/.ollama. On a plain Linux install the FAQ gives /usr/share/ollama/.ollama/models, changeable with OLLAMA_MODELS.