In this article

Your laptop may run a small local AI model, but installed RAM alone cannot tell you which model will be useful. Model size, quantization, context length, GPU memory and the rest of your workload all affect the result. Try a modest workload before buying hardware.

This guide explains the decision process. We have not benchmarked the laptops or models discussed here; the arithmetic below is illustrative.

Separate storage from working memory#

A model download occupies disk space. Running it also needs working memory for its weights, runtime overhead and the conversation or document context. A model file that fits on your SSD may still be impractical to run.

A rough weight-only calculation for a hypothetical eight-billion-parameter model stored at four bits per parameter is:

text
8 billion parameters × 4 bits ÷ 8 = 4 billion bytes

That is not the full runtime requirement. Quantization metadata, caches and other allocations add overhead. Do not interpret this example as a four-gigabyte hardware recommendation.

Context length changes the workload#

A short question and a long document can use the same model but require different memory. Longer context increases memory requirements, as Ollama's context documentation explains.

Start with the context you actually need. If you mostly rewrite short paragraphs, buying for a huge document window may be unnecessary. If you need to analyze long files, verify that the model can handle the relevant content and that your hardware can sustain the workload.

Understand RAM, VRAM and unified memory#

Dedicated graphics memory is not identical to system RAM. Some systems use a unified memory design; others split memory between CPU and GPU.

Ollama's FAQ documents ollama ps, which reports loaded models and their CPU/GPU placement. That observation is more useful than assuming the GPU is doing all the work because your laptop has one.

When a workload falls back to a slower execution path, adding system RAM may improve capacity without delivering the responsiveness you expected. Check both whether the model fits and whether its speed is acceptable.

Run a suitability trial#

Use the following procedure with non-sensitive sample content:

  1. Install a supported runtime from its official source.
  2. Choose a small model whose license permits your intended use.
  3. Run a short task similar to your real work.
  4. Observe memory pressure, model placement and responsiveness.
  5. Increase the document length gradually.
  6. Repeat with your normal browser and other applications open.

Score usefulness separately from speed. A fast answer that invents facts is not a successful trial. Use the AI answer checking guide to evaluate important claims.

Local execution is a boundary you must verify#

Ollama distinguishes local execution from cloud-hosted models and provides controls for disabling cloud features in its FAQ. A desktop interface does not automatically mean every request stays on your device.

Check the selected model, integrations and extensions. Web search, synchronization or a remote document connector can introduce additional data flows. Keep an inference service local unless you deliberately configure authenticated network access.

Decide what to buy—or whether to buy anything#

For occasional experiments, use the machine you own first. For a frequently used workload, choose hardware around the tested model and required context. Consider battery use, noise, storage and upgradeability alongside capacity.

Compare the full cost with a hosted service, including the time spent maintaining the local setup. Local execution can be valuable for control and offline work, but those benefits do not guarantee better answer quality.

If you are choosing a general-purpose laptop rather than a dedicated inference machine, see 16GB versus 32GB RAM. The right configuration starts with the workload, not a universal number.