Independent Australian Storage & Infrastructure Authority

Home › Tools › AI Hardware Requirements Calculator

AI Hardware Requirements Calculator

This calculator works out how much memory a local AI model needs on a NAS or home server, starting from Ollama's own download size for the model, then shows which NAS models hold it and what running it costs against a cloud API at your own token use.

Pick a model, the context length and how many requests run at once. The memory figure is the model's size plus an allowance for the context and memory kept free for the NAS; both are NTK planning figures you can change, or enter what ollama ps shows for your model.

1. The model

Filled from Ollama's download size when you pick a model. For another model, use the size on its Ollama page.
How much text the model keeps in view. Agents and coding tools need much more than a chat.
People or apps asking at once. Each one adds context memory, not another copy of the model.
Planning figures and a measured size
2 GB is an NTK planning default for the NAS system and its other apps.
15% is an NTK planning figure, not Ollama's: Ollama publishes no per model figure.
If you have run the model, its SIZE in ollama ps replaces the model size and the allowance.

Memory needed

Memory to plan for
Cannot calculate
NAS in NTK's list that hold it as shipped
Cannot calculate
Speed
No figure
On a NAS, Ollama runs on the processor. No primary benchmark exists for these processors, so NTK gives no tokens per second figure.

NAS models that can hold it

NASMemoryResultMarket price

Own an older NAS? Check it in Can My NAS Run AI?, which also covers Immich, Synology Photos and Whisper.

2. Cloud API or local: what it costs

A starting figure: replace it with yours from your API usage page.
Output tokens cost more than input on every model listed.
Needed only when the NAS ships with less memory than the model needs.
Filled from the maker's busy figure where NTK has one. If you already own the NAS, enter only the extra watts AI adds (a plug-in meter shows it).
24 = that figure all day. With the maker's busy figure, a NAS that is idle most of the day will usually use less.
Cents per kWh from your bill, or your state's regulated or reference rate.

Cost comparison

Enter your usage above to see the cost comparison.

Cloud API a year
Cannot calculate
Local: upfront
Cannot calculate
Local: electricity a year
Cannot calculate

A local model and a cloud model are not equally capable; this compares cost only. Drives, your time and the NAS's other jobs are not counted.

Every cloud model at your usage

ModelA monthA year

How the Calculator Works

The model's size. Ollama publishes the download size of each model it offers, at a stated quantisation (how many bits each weight is stored in). That size is the starting point: the whole model has to sit in memory.

Context. Ollama's FAQ says it uses a context window of 4,096 tokens by default. Ollama's documentation says a larger context length increases the memory needed to run a model, and that tasks needing a large context, such as agents and coding tools, should be set to at least 64,000 tokens. Ollama's FAQ says parallel requests to one model multiply its context size, and the memory needed scales with the number of parallel requests times the context length. Ollama's FAQ says the ollama ps command shows the models loaded into memory with their size, and 100% CPU means the model was loaded entirely into system memory. Ollama's hardware page lists NVIDIA GPUs (compute capability 5.0 and newer), AMD GPUs through ROCm, Apple GPUs through Metal, and further GPUs on Windows and Linux through Vulkan, which on most Linux distributions needs extra components installed. It does not list NPUs. Ollama publishes no per model figure, so NTK adds a planning allowance of 15% of the model size for every 4,096 tokens held in memory (the context length times the requests running at once). It is labelled as NTK's, and you can change it.

The NAS itself. 2 GB is kept free for the NAS system and its other apps, an NTK planning default you can change.

Measured beats planned. If you have run the model, the SIZE that ollama ps shows replaces the model size and the allowance.

Which NAS hold it. The total is compared with each NAS's memory as shipped and its maker's maximum, from NTK's NAS specification database. No speed figure is given: no primary benchmark of Ollama on these NAS processors exists.

Cost. Cloud = your input and output tokens a day times each vendor's published price per million, in US dollars, converted with the RBA rate (1.4333 Australian dollars per US dollar, 2026-09-30). Local = the hardware you buy for this (nothing if you own it), any memory upgrade, and electricity for the hours it runs at your rate. Picking a NAS fills in its maker's busy figure; at 24 hours a day the NAS is counted at that figure all day. Makers measure it during file transfers or typical operation, and Synology says its figures are not meant to be minimum or maximum power consumption figures (Synology, checked 2026-10-14), so a NAS that is idle most of the day will usually use less. A figure you measured at the wall is better. Break-even is shown only when it exists, and never as negative years.

Memory to plan for, by model

Model (Ollama)ParametersQuantisationDownloadContext windowMemory to plan forSource
Llama 3.2 3B (llama3.2:3b)3.21BQ4_K_M2 GB128K tokens4.3 GBOllama, checked 2026-10-18
Qwen3.5 4B (qwen3.5:4b)4.21BQ4_K_M3.3 GB256K tokens5.8 GBOllama, checked 2026-10-18
Llama 3.1 8B (llama3.1:8b)8.03BQ4_K_M4.9 GB128K tokens7.7 GBOllama, checked 2026-10-18
Qwen3 8B (qwen3:8b)8.19BQ4_K_M5.2 GB40K tokens8 GBOllama, checked 2026-10-18
DeepSeek-R1 8B (deepseek-r1:8b)8.19BQ4_K_M5.2 GB128K tokens8 GBOllama, checked 2026-10-18
Gemma 4 E4B (gemma4:e4b)7.52BQ4_K_M6.6 GB128K tokens9.6 GBOllama, checked 2026-10-18
Qwen3.5 9B (qwen3.5:9b)8.95BQ4_K_M6.6 GB256K tokens9.6 GBOllama, checked 2026-10-18
Gemma 4 12B (gemma4:12b)11.9BQ4_K_M8 GB256K tokens11.2 GBOllama, checked 2026-10-18
Phi-4 14B (phi4:14b)14.7BQ4_K_M9.1 GB16K tokens12.5 GBOllama, checked 2026-10-18
Qwen3 14B (qwen3:14b)14.8BQ4_K_M9.3 GB40K tokens12.7 GBOllama, checked 2026-10-18
gpt-oss 20B (gpt-oss:20b)20.9BMXFP414 GB128K tokens18.1 GBOllama, checked 2026-10-18
Gemma 3 27B (gemma3:27b)27BQ4_K_M17 GB128K tokens21.6 GBOllama, checked 2026-10-18
Qwen3.5 27B (qwen3.5:27b)26.9BQ4_K_M17 GB256K tokens21.6 GBOllama, checked 2026-10-18
Qwen3 30B (mixture of experts) (qwen3:30b)30.5BQ4_K_M19 GB256K tokens23.9 GBOllama, checked 2026-10-18
Llama 3.1 70B (llama3.1:70b)70.6BQ4_K_M43 GB128K tokens51.5 GBOllama, checked 2026-10-18

Memory to plan for = download + the NTK allowance at Ollama's default 4,096 token context and one request + 2 GB kept free for the NAS, rounded up.

{# s268 T543 (2): the "A month at " column was dropped: a monthly cost for named models at a usage NTK chose is a derived figure in tool copy (s263 rule, s264 scope). The reader's own monthly cost is the "Every cloud model at your usage" table. #}

Cloud API prices

ModelInput, per million tokensOutput, per million tokensSource
OpenAI GPT-6 Astra
Standard processing, short context (requests up to 272K input tokens)
Short context rate. OpenAI prices a request with more than 272K input tokens at 2x input and 1.5x output for the whole request.		US$10US$50OpenAI, checked 2026-10-18
Currency not named on the page; treated as US dollars
OpenAI GPT-6.1 Sol
Standard processing, short context (requests up to 272K input tokens)
Short context rate. OpenAI prices a request with more than 272K input tokens at 2x input and 1.5x output for the whole request.		US$2US$10OpenAI, checked 2026-10-18
Currency not named on the page; treated as US dollars
OpenAI GPT-6 Luna
Standard processing, short context (requests up to 272K input tokens)
Short context rate. OpenAI prices a request with more than 272K input tokens at 2x input and 1.5x output for the whole request.		US$0.10US$0.50OpenAI, checked 2026-10-18
Currency not named on the page; treated as US dollars
Anthropic Claude Opus 5.5
Base input and output tokens
One rate at any request length: Anthropic includes the full 1M token context window at its standard price.		US$4US$20Anthropic, checked 2026-10-18
Anthropic Claude Sonnet 5.5
Base input and output tokens
One rate at any request length: Anthropic includes the full 1M token context window at its standard price.		US$2US$10Anthropic, checked 2026-10-18
Anthropic Claude Haiku 5.5
Prompts up to 100,000 tokens (Anthropic charges more above that)
Rate for prompts up to 100,000 tokens. Anthropic charges higher prices for a request whose prompt is over 100,000 tokens.		US$0.10US$0.50Anthropic, checked 2026-10-18
Google Gemini 3.6 Flash
Paid tier, standard. Google's price through December 31, 2026. Google lists US$1.50 in and US$7.50 out per million tokens starting January 1, 2027.
Google lists one rate for this model, with no request length tier.		US$0.75US$3.75Google, checked 2026-10-18
Google Gemini 3.5 Flash-Lite
Paid tier, standard, text input
Google lists one rate for this model, with no request length tier.		US$0.30US$2.50Google, checked 2026-10-18

Privacy

Local inference keeps prompts and documents on hardware you control. If you handle personal information, the OAIC explains your obligations (oaic.gov.au).

Buying

Prices come from NTK's market price index of Australian retailers (2026-10-23); a NAS with too few current listings shows no price. Your first step for a faulty product is usually the retailer that sold it. The ACCC explains your rights under the Australian Consumer Law at accc.gov.au.

Assumptions NTK made

    Method last reviewed: 2026-10-14. Model sizes and cloud prices last checked at the source: 2026-10-18. Exchange rate: RBA statistical table F11 (Exchange Rates), series FXRUSD, WM/Reuters, 2026-09-30.

    Frequently Asked Questions

    How much memory does a local AI model need on a NAS?
    Start from the model's own size, which Ollama publishes for every download: llama3.1:8b is 4.9 GB (Q4_K_M quantisation). Add memory for the context: NTK plans 15% of the model size for every 4,096 tokens held in memory, about 0.8 GB at Ollama's default context, an NTK planning figure rather than Ollama's. Keep 2 GB free for the NAS itself, also an NTK planning default. That is 7.7 GB in total, which 19 NAS models in the calculator hold as shipped. For a real figure, run the model and read the SIZE that ollama ps shows.
    What do context length and parallel requests change?
    Ollama's documentation says a larger context length increases the memory needed to run a model, and that tasks needing a large context, such as agents and coding tools, should be set to at least 64,000 tokens. Ollama's FAQ says parallel requests to one model multiply its context size, and the memory needed scales with the number of parallel requests times the context length. In the calculator the allowance grows with the context length times the parallel requests: llama3.1:8b at 32,768 tokens needs about 12.8 GB, and four parallel requests at the default context about 9.9 GB. The model itself is loaded once.
    What does quantisation change?
    Each Ollama download is one quantisation of a model. The calculator uses the default download Ollama lists for each model (MXFP4 and Q4_K_M). A version with fewer bits is a smaller download and needs less memory, and it can change the answers; how much depends on the model and the task, so NTK gives no quality percentage. To size another version, choose Other and enter the size from its Ollama page.
    Which models does the calculator include?
    15 current models from Ollama's library, small to large: llama3.2:3b, qwen3.5:4b, llama3.1:8b, qwen3:8b, deepseek-r1:8b, gemma4:e4b, qwen3.5:9b, gemma4:12b, phi4:14b, qwen3:14b, gpt-oss:20b, gemma3:27b, qwen3.5:27b, qwen3:30b and llama3.1:70b. Each size, quantisation and context window is taken from the model's Ollama page and dated (checked 2026-10-18). Any other model can be sized by entering its size.
    Is running AI on a NAS cheaper than a cloud API?
    It depends on your token use, the cloud model and what you would buy, so the calculator works it out for your figures. In round numbers (not a real model or plan): if the cloud API would cost $500 a year at your usage, and a NAS bought for this costs $1,000 and draws 30 W all day, its electricity is about 263 kWh a year. Over the years you count, add the cloud bill each year against the NAS price plus its electricity at your rate each year, and see which side is lower. Makers measure a NAS's busy figure during file transfers or typical operation, and Synology says its figures are not meant to be minimum or maximum power consumption figures, so a NAS that sits idle most of the day will usually use less. The local model is not the same model as the cloud one, so this compares cost only.
    Where do the cloud prices come from?
    From each vendor's own pricing page: Anthropic, Google and OpenAI, per million input tokens and per million output tokens at standard rates, checked 2026-10-18. Anthropic's pricing page says its prices are in US dollars. Google's Gemini API pricing page gives paid tier prices per million tokens in US dollars. OpenAI's pricing page shows prices with a $ sign and does not name the currency; OpenAI's API documentation says spend alert amounts must be in USD, so NTK converts OpenAI's prices as US dollars. Prices are converted to Australian dollars with the RBA monthly rate NTK stores (1.4333, 2026-09-30). API use is billed separately from chat subscriptions.
    Can a NAS's GPU or NPU speed up Ollama?
    Ollama's hardware page lists NVIDIA GPUs (compute capability 5.0 and newer), AMD GPUs through ROCm, Apple GPUs through Metal, and further GPUs on Windows and Linux through Vulkan, which on most Linux distributions needs extra components installed. It does not list NPUs. NTK found no NAS maker documenting Ollama GPU acceleration on its NAS, so the calculator treats a NAS as running Ollama on its CPU and gives no speed figure. Other apps differ: Immich documents GPU acceleration for its machine learning through ARM NN (Mali), CUDA (NVIDIA), ROCm (AMD), OpenVINO (Intel GPUs such as Iris Xe and Arc) and RKNN (Rockchip), and calls the feature experimental.
    Does running AI locally help with privacy?
    Local inference keeps prompts and documents on hardware you control. If you handle personal information, the OAIC explains your obligations (oaic.gov.au).
    Can a NAS run image generation?
    ComfyUI's README says it supports NVIDIA, AMD, Intel, Apple Silicon and Ascend GPUs, and that its portable Windows build can also run on the CPU only. NTK found no speed figure for image generation on NAS processors, so it gives none, and this calculator covers language models.

    Related Tools & Guides

    Can My NAS Run AI? → Best NAS for AI in Australia → LLM Quantisation Explained → Mini PC vs NAS for Local AI → NAS Power Cost Calculator → All Tools →