Skip to content
LandGadget
Computers & AIBuying Guides

Buying a laptop for local AI: how much memory, GPU and NPU you really need

Want to run AI models on your own laptop? This guide explains how model size maps to memory, when an NPU helps, and which specs to prioritise.

LandGadget Staff · · 2 min read

Share
A backlit laptop keyboard
Image: Colin / Wikimedia Commons, CC BY-SA 4.0
On this page

Running AI models on your own laptop keeps your data private, works offline and costs nothing per prompt. The specs that matter for this are not the usual ones.

Memory is the main limit

A model has to fit in memory. A rough rule for 4-bit quantised models, the usual format for local use, is about 0.6 GB per billion parameters, plus room for context:

Model sizeMemory needed (approx.)Good for
3–4B4–6 GBquick chat, summaries
7–8B6–8 GBeveryday assistant, coding help
13–14B10–12 GBbetter reasoning
30–32B20–24 GBstrong all-rounder
70B40–48 GBnear cloud quality

Two ways to get that memory

  • Discrete GPU (NVIDIA RTX): very fast, but limited by VRAM. For example, 8 GB on an RTX 5060 laptop and 16–24 GB on top models.
  • Unified memory (Apple M-series, AMD Ryzen AI Max): the CPU and GPU share one large pool. Speed per token is lower than a top GPU, but you can load much larger models. A 64–128 GB unified machine runs 70B models that no laptop GPU can.

Where the NPU fits

NPUs, such as those in Copilot+ PCs rated at 40+ TOPS, are efficient for small always-on tasks: background blur, live captions and small on-device models. For big local language models, most tools today still use the GPU.

What to buy

  • Curious beginner: 16 GB RAM is enough for 7–8B models.
  • Serious hobbyist or developer: 32 GB unified memory, or an RTX GPU with 12 GB or more of VRAM.
  • Big models: 64–128 GB of unified memory.

Also check that you have fast SSD storage of at least 1 TB, because each model file is several gigabytes, and good cooling, because long generations push laptops hard. Free tools such as LM Studio or Ollama will get you running in minutes.

Share

Keep reading