Why run AI on your own machine, and what you give up

What you will be able to do

• Explain what local processing can keep on your device, and which settings and tools still need checking.
• Name what you give up compared with frontier cloud models, in concrete terms.
• Place yourself on the decision this path builds toward: use what you have, buy a device, or build a PC.

The promise, stated precisely

A local model runs on a computer you control. Once the model file is on the machine, the question, the documents, and the answer can all stay there: nothing needs to travel to a provider's servers. That is the whole promise. It replaces "read the provider's terms and trust them" with "check what leaves the machine", which is something you can test yourself (lesson 9.9 shows how).

It is narrower than it sounds. Local does not mean secure: the machine still needs disk encryption, a login, backups, and the firm's usual controls. Local does not mean offline by default: many apps check for updates, download models, or offer optional cloud features. Ollama, for example, documents a setting to turn its cloud features off entirely. [source] And local does not mean accurate: a model on your desk invents cases as readily as one in a data centre.

What you give up

The largest frontier models run on clusters of data-centre accelerators with hundreds of gigabytes of fast memory. The open-weight models you can run at home are smaller. In October 2026 the practical sizes for a single machine range from about 9 billion parameters on an ordinary laptop to roughly 120 billion on a large unified-memory machine; OpenAI's gpt-oss-120b, for instance, has 117 billion parameters and is designed to fit a single 80 GB data-centre GPU. [source]

In practice that means four trade-offs. Capability: smaller models make more reasoning errors on long, tangled problems. Speed: answers stream more slowly unless the hardware is well chosen. Convenience: no built-in web search, no polished document pipeline unless you add one. Maintenance: you update the software and choose the models; nobody does it for you.

What you keep

You keep custody. For first-pass extraction, chronologies, summaries of a single record, redaction checks, and drafting from supplied material, a model may help, but test its results and all routes by which the app can send data. The professional duties are the same either way: Confidentiality and supervision of the output still apply, and nothing here changes the need to verify every citation.

The minimum to try it today

LM Studio states its own floor: on a Mac, Apple Silicon with macOS 14 or newer and 16 GB of unified memory recommended; on Windows, a processor with AVX2 support, 16 GB of RAM, and at least 4 GB of dedicated graphics memory recommended. [source] If your machine meets that, you can run a small model this afternoon and decide from experience whether the bigger investment is worth it.

Self-check

1. What can local processing change, and what still needs checking?
2. Name two tasks where a mid-size local model is often good enough, and one where a frontier cloud model is still likely to be better.
3. What are the three routes this path helps you choose between?
4. Does a local model reduce the need to verify its citations?