Cloud or local: where should your AI run in 2026?

"Cloud" means the model runs on someone else's servers: you type into ChatGPT, Gemini or Claude, your words travel over the internet, and the answer comes back. "Local" means the model runs on your own computer, using a free tool such as Ollama or LM Studio and an open-weight model, one whose trained numbers anyone can download. Both are mature choices in early 2026, and the line between them is blurring: Ollama, best known for local models, has offered optional cloud models since September 2025, run with the same commands.

So the question is rarely "which is better?" It is "which is better for this task, with this data, on this machine?" Four factors settle it.

The four factors

Privacy. With a local model, your prompt is processed on your own hardware. Ollama's documentation states plainly that local use keeps your prompts and data out of its view. With a cloud service, your text is sent to the provider, and what happens next depends on its policy and your settings.

Quality. The largest and strongest models are only available in the cloud, or need server hardware to run. A laptop typically runs models in the range of a few billion to about 30 billion parameters (the adjustable numbers a model learns). They are good at drafting, summarizing and rewording, and weaker on hard reasoning, long documents and up-to-date facts.

Cost. Local is free per question once you own the computer, but a capable machine with lots of memory is not cheap. Cloud is cheap to start, with free tiers, and then costs a monthly subscription or a fee per use.

Maintenance. A cloud chatbot needs nothing from you. Local means installing software, downloading multi-gigabyte models, updating them and choosing sizes that fit your memory.

Three worked examples

These are illustrative profiles, not real people, to show how the table plays out.

A family doctor's office manager wants help rewording letters that mention patients. The text includes health details, so privacy wins. A small local model on the office computer handles rewording well. Hard questions with no patient data, such as summarizing a new public guideline, can still go to a cloud chatbot.

A retired engineer writing a local history book wants feedback on long chapters and help checking facts. Nothing is confidential, and the chapters are long. A cloud model with a large context window (how much text it can read at once) is the better fit. A local model would be slower and would struggle with the length.

A small web agency wants AI inside its own tools and is watching costs. It can prototype locally for free with Ollama, then send only the heavy jobs to a hosted model through an API. With Ollama, the same command line can do both: a model name ending in -cloud runs on Ollama's servers once you sign in.

ollama run llama3.2              # runs on your computer
ollama run gpt-oss:120b-cloud    # runs on Ollama's servers (sign-in required)

That one suffix is easy to miss, and it changes where your words go. Ollama says its cloud does not keep prompts or train on them, but the text still leaves your machine.

A simple rule of thumb

Start from the data, not the tool. If the text contains someone else's personal information, client material, or anything under a confidentiality agreement, either run it locally or use a cloud plan whose contract covers that kind of data. If it does not, choose on quality and convenience, which usually means the cloud. Keep a small local model installed anyway: it is a useful fallback when you are offline and a safe place to experiment.

What does privacy really require?

Privacy is not a property of a brand; it is a chain, and every link has to hold. Ask three questions of any setup. Where is the text processed? Is it stored, and for how long? Who can read it, including people at the provider?

Cloud services answer these in their privacy notices, and the answers can surprise you. Google's Gemini notice, for example, has warned users since 2024 not to enter confidential information they would not want a human reviewer to see, and says that reviewed conversations can be kept for up to three years, separately from your account, even after you delete your activity. Settings that limit training or history help, but read what each one actually does.

Local is not automatically private either. By default, Ollama's server accepts connections only from your own computer (on port 11434); if you change that to share it on your network, anyone on that network may reach it. Files you feed a local model stay on your disk, along with any chat logs your app saves. And a "local" app that calls a cloud model, like the -cloud example above, is no longer local for that request.

In practice, privacy requires three habits: know which mode each tool is in, keep sensitive material on tools whose data path you have checked, and remove names and identifying details when a task can be done without them. A prompt that never contained the private detail cannot leak it.