AI models you can run on a NAS
38 open models across 6 families, with their size at each quantization from the Ollama library and the NAS memory each one needs.
Llama
Llama 3.1 8B runs on any AI NAS with 16GB of memory, and Llama 3.3 70B needs around 60GB at 4-bit (Estimated). Llama 4 Scout and Maverick are mixture-of-experts models that need far more memory than most NAS systems support.
7 models from 1.24B to 402B
Qwen
Qwen3 8B runs on any AI NAS with 16GB of memory, Qwen3 32B needs about 32GB, and Qwen2.5 72B needs around 65GB at 4-bit (Estimated). Qwen3 30B-A3B is a mixture-of-experts model that fits in about 31GB and answers faster than its size suggests.
12 models from 0.75B to 235B
DeepSeek
The DeepSeek-R1 distilled models run well on an AI NAS: the 8B needs about 15GB and the 32B about 32GB of memory (Estimated). The full DeepSeek-R1 and DeepSeek-V3 671B models need hundreds of gigabytes and are beyond any desktop NAS.
9 models from 1.78B to 671B
Gemma
Gemma 3 4B runs on almost any AI NAS, the 12B needs about 18GB of memory and the 27B about 29GB at 4-bit (Estimated). The 4B, 12B and 27B models also understand images, which suits photo and document tasks on a NAS.
4 models from 1B to 27B
Mistral
Mistral 7B and NeMo 12B run on any AI NAS with 16GB to 24GB of memory. Mistral Small 24B needs about 25GB, and Mistral Large 123B needs around 96GB at 4-bit (Estimated).
4 models from 7.25B to 123B
Phi
Phi-4-mini, at 3.8B parameters, runs on almost any AI NAS with about 11GB of memory. Phi-4 14B needs about 19GB at 4-bit (Estimated) and gives strong reasoning for its size, which suits NAS systems without a GPU.
2 models from 3.84B to 14.7B
Check whether your NAS can run a model · Work out how much memory you need