Use a cloud chat app (ChatGPT, Claude, Gemini) if you want the most capable AI with zero setup and you are comfortable that your conversations are processed on a company's servers. Run a model locally if privacy, offline use, no subscription, or hands-on control matter more than raw capability. Most people should use a cloud chat app. The privacy-conscious, the offline, and the tinkerers should run local. Many people sensibly do both: a cloud app for hard tasks, a local model for private or routine ones.
This is the decision between the two practical ways an ordinary person uses an LLM: a hosted cloud chat app where the company runs the model for you, or a local model running on your own machine.
The developer API is a third path covered in how-to-use-an-llm, but for a non-coder the real choice is cloud versus local. The two are not rivals so much as tools for different priorities, and the right answer depends entirely on what you value most.
Choosing on instinct usually leads to a poor fit. People who would be perfectly happy with a free cloud app sometimes spend a weekend setting up local AI they do not need. People who must protect sensitive data sometimes paste it into a cloud app without realising it left their control. Naming the tradeoffs plainly (capability, privacy, cost, setup, offline) turns a vague preference into a clear decision, and shows when running both is the smart move rather than picking one.
| What you care about | Cloud chat app | Local model |
|---|---|---|
| Capability | Highest. The frontier models. | Good but lower. Small and mid-size open models. |
| Setup | None. Open a browser or app. | Install a program, download a model. Easy now, but a step. |
| Privacy | Conversations go to the company, and may train its models unless you opt out. | Strongest. Nothing leaves your computer. |
| Cost | Free tier, or about 20 US dollars a month for the best. | Free to run. Cost is the hardware you own and electricity. |
| Offline | No. Needs an internet connection. | Yes. Works with no internet once downloaded. |
| Up-to-date info | Often, via built-in web search. | No live web access by default, and a fixed knowledge cutoff. |
| Hardware needed | Any device with a browser. | A reasonably capable PC or Mac. See hardware-and-performance. |
| Control | The provider can change or retire the model. | You keep the exact model forever. |
This describes most people most of the time.
These are not mutually exclusive, and combining them is often the best setup. A common pattern is a cloud subscription for the genuinely hard problems, plus a local model for private documents and routine drafting that does not need the frontier. You get top capability when it matters and full privacy when it counts, without overpaying or oversharing.
| Term | Plain meaning |
|---|---|
| Cloud / hosted | The model runs on a company's servers. You reach it over the internet. |
| Local / on-device | The model runs on your own computer, offline, with nothing sent out. |
| Frontier model | The most capable models available, run in the cloud by the big providers. |
| Open-weight model | A downloadable model you can run locally. Smaller and weaker than the cloud frontier. |
| Knowledge cutoff | The date after which a model knows nothing, unless it can search the web. See training-vs-inference. |
| Opt out of training | A setting in many cloud apps that stops your conversations being used to improve the model. |
"Local is the private and safe choice, so it is simply better."
Only if its lower capability is acceptable and you have the hardware. For a hard reasoning task, a frontier cloud model may give a far better answer. Privacy is one axis, not the whole picture.
"Cloud apps are unsafe with my data."
They are not inherently unsafe, but your data does go to the company. Whether that is acceptable depends on the data and the provider's terms. For ordinary use it is usually fine. For confidential material, prefer local or a business tier with a no-retention guarantee.
"Paying for cloud means I no longer need anything local."
Plenty of people keep a local model specifically for private or offline work even while paying for a cloud app. The two solve different problems.
"Local models are free, so they cost nothing."
The software and models are free, but a capable local setup needs adequate hardware, and heavy local use draws power and stresses a laptop. The cost moved, it did not vanish.
"I have to pick one forever."
You do not. Trying a free cloud app and a free local model side by side, then keeping whatever fits each task, is the most sensible approach.
Every claim traces to a cited source below.