AI models you can actually run on a NAS
Chat, reasoning, coding, vision and embedding models — with the size, RAM and GPU each one needs, so you can match a model to the NAS you already have. Quantized download sizes come from Ollama's public model library. Facts as of 2026-11.
Chat / reasoning / coding / vision models
| Model | Params | Download (Q4) | Min RAM | Comfort RAM | GPU/VRAM | What it's for |
|---|---|---|---|---|---|---|
| Llama 3.2 3B Meta Llama | 3B | 2 GB | 4 GB | 8 GB | 4 GB VRAM CPU-only works, GPU = faster | 轻量聊天/摘要，几乎任何 x86 NAS 都够 |
| Llama 3.1 8B Meta Llama | 8B | 4.9 GB | 8 GB | 16 GB | 8 GB VRAM CPU-only works, GPU = faster | 8B 主力档，16GB 才舒服；8GB 默认机是下限 |
| Llama 3.3 70B Meta Llama | 70B | 43 GB | 64 GB | 128 GB | 48 GB VRAM CPU-only works, GPU = faster | 旗舰档，NAS 上要 64-128GB 内存，基本需加内存或 M.2 卡 |
| Qwen 3 8B Alibaba Qwen | 8B | 5.2 GB | 8 GB | 16 GB | 8 GB VRAM CPU-only works, GPU = faster | 中英均衡，NAS 上最流行的 8B 之一 |
| Qwen 2.5 14B Alibaba Qwen | 14B | 9 GB | 16 GB | 32 GB | 16 GB VRAM CPU-only works, GPU = faster | 14B 档，16GB 紧、32GB 舒服 |
| Qwen 2.5 32B Alibaba Qwen | 32B | 20 GB | 32 GB | 64 GB | 24 GB VRAM CPU-only works, GPU = faster | 32B 档，需要 32GB 起步的机器 |
| Qwen 2.5 72B Alibaba Qwen | 72B | 47 GB | 64 GB | 128 GB | 48 GB VRAM CPU-only works, GPU = faster | 72B 旗舰，NAS 上要 64GB 以上 |
| Mistral 7B Mistral AI | 7B | 4.4 GB | 8 GB | 16 GB | 8 GB VRAM CPU-only works, GPU = faster | 经典 7B，与 8B 同档 |
| Mixtral 8x7B Mistral AI | 46B MoE | 26 GB | 32 GB | 64 GB | 32 GB VRAM CPU-only works, GPU = faster | MoE 只激活 14B 但文件 26GB，需要 32GB+ 装下 |
| GPT-OSS 20B OpenAI | 21B MoE | 14 GB | 32 GB | 48 GB | 24 GB VRAM CPU-only works, GPU = faster | MoE 小激活，14GB 文件，32GB 舒服 |
| Phi 4 14B Microsoft | 14B | 9.1 GB | 16 GB | 32 GB | 16 GB VRAM CPU-only works, GPU = faster | 小身材高推理，14B 档 |
| DeepSeek R1 8B DeepSeek | 8B | 5.2 GB | 8 GB | 16 GB | 8 GB VRAM CPU-only works, GPU = faster | 8B 推理模型，思考过程吃显存，16GB 更稳 |
| Qwen 2.5 Coder 32B Alibaba Qwen | 32B | 20 GB | 32 GB | 64 GB | 24 GB VRAM CPU-only works, GPU = faster | 编码档，32GB 起步；NAS 上跑代码助手 |
| Qwen 3 Coder 30B Alibaba Qwen | 30B MoE | 19 GB | 32 GB | 64 GB | 24 GB VRAM CPU-only works, GPU = faster | MoE 编码模型，19GB 文件 |
| DeepSeek Coder V2 16B DeepSeek | 16B MoE | 8.9 GB | 16 GB | 32 GB | 16 GB VRAM CPU-only works, GPU = faster | MoE 编码小档，16GB 可跑 |
| LLaVA 7B LLaVA | 7B | 4.7 GB | 8 GB | 16 GB | 8 GB VRAM CPU-only works, GPU = faster | 视觉语言模型，8B 档；NAS 相册问答用 |
| Qwen 2.5 VL 7B Alibaba Qwen | 7B | 6 GB | 8 GB | 16 GB | 8 GB VRAM CPU-only works, GPU = faster | 视觉语言，中文识别强 |
| MiniCPM-V OpenBMB | 8B | 5.5 GB | 8 GB | 16 GB | 8 GB VRAM CPU-only works, GPU = faster | 端侧视觉，体积小 |
| Nomic Embed Nomic | 0.12B | 0.27 GB | 2 GB | 4 GB | 1 GB VRAM CPU-only works, GPU = faster | RAG 嵌入模型，任何 NAS 都跑得动；检索用不是聊天 |
| BGE-M3 BAAI BGE | 0.6B | 1.2 GB | 2 GB | 4 GB | 2 GB VRAM CPU-only works, GPU = faster | 多语言嵌入，NAS 知识库检索 |
"Min RAM" fits the model with OS + file services still running. "Comfort RAM" gives headroom for context and concurrency. GPU/VRAM is the size that makes it feel fast; CPU-only always works, just slower. See which NAS models reach these RAM levels →
Embedding & RAG models (the cheapest win)
These power "search your own data" — turn documents, photos and notes into vectors so a chat model can quote them. They are tiny and run on any NAS.
| Model | Download | Min RAM | Use |
|---|---|---|---|
| Nomic Embed Nomic | 0.27 GB | 2 GB | RAG 嵌入模型，任何 NAS 都跑得动；检索用不是聊天 |
| BGE-M3 BAAI BGE | 1.2 GB | 2 GB | 多语言嵌入，NAS 知识库检索 |
How to read this against your NAS
Default 4-8 GB NAS → 3B chat, 7B at a stretch, any embedding model. Plenty of home automation and summarization, no 14B.
16 GB (QNAP TS-464 upgraded, UGREEN i5) → 7-8B comfortable, 14B tight. This is where most "AI NAS" marketing points.
32-64 GB (QNAP 64 GB SKUs, UGREEN DDR5, DIY) → 14B/32B, MoE models like GPT-OSS and Mixtral that only activate a small slice.
Above 64 GB / with a dGPU → 70B class. On closed NAS this means the QNAP TVS-AIh1688 or a DIY fnOS build — see the AI-NAS reality check for which boxes are actually built for it.
Not sure which model fits your box? Use the AI NAS fit checker — pick a model, pick your NAS, get a verdict.
FAQ
What is the best AI model to run on a NAS?
For most NAS RAM budgets, 7-8B models (Llama 3.1 8B, Qwen 3 8B, Mistral 7B) are the sweet spot: they fit in 8-16 GB and give usable chat quality. If you can upgrade to 32-64 GB, 14-32B models (Qwen 2.5 32B) are a big step up. Above that, a DIY build is the realistic route.
Does an AI model on a NAS need a GPU?
No — any model runs on CPU-only. A GPU (or the iGPU in a QNAP/UGREEN i5/Intel box) makes the same model many times faster. So a GPU is a speed upgrade, not a requirement. For 8B-class models, CPU-only is slow-but-usable; a GPU makes it feel instant.
What are embedding models good for on a NAS?
Embedding models (Nomic Embed, BGE-M3) are tiny — under 1.5 GB — and power RAG: they turn your documents, photos or notes into vectors so a chat model can search and quote your own data. They are the cheapest win on any NAS and need almost no RAM.
Are these numbers accurate?
Quantized download sizes are from the public Ollama model library (the default 4-bit download). Total RAM adds operating headroom and KV-cache. These are planning numbers, not marketing specs; verify on your exact build as of Nov 2026.