Run AI on Your Own Laptop: The 16GB Guide

01 What local AI actually is

In May 2025, a U.S. judge ordered OpenAI to keep ChatGPT conversations it would normally delete. The order ended that September. The question didn't: once your file is in someone else's cloud, who decides what happens to it? Running the AI on your own laptop is the third option.

A model is a file. A big one, but a file. It reads your text and writes the answer. A local app runs that file on your own chip. Two free ones: LM Studio, a desktop app with a chat window, and Ollama, which you can run from your computer's terminal. Nothing goes out. Ollama's privacy policy says your data stays on your machine. LM Studio's docs say it runs entirely offline once you have the model files.

One caveat. Ollama also offers cloud-hosted models, and prompts sent to those are processed on its service. Local means the models you download.

02 The 16GB tray

Since October 2024, Apple's base MacBook Air has come with 16GB of memory. Picture that memory as a tray. Three things share it:

Your operating system and open apps. They take a corner. The model. Its size is fixed. The conversation. Every message, pasted page and answer stays in memory. Google says the model sizes it quotes leave this out: the conversation needs room on top.

Model in Ollama Download size On 16GB

Qwen3 8B, the smaller version 5.2GB Fits, with room to spare

Qwen3 30B, the bigger version 19GB Doesn't fit at all

A one-page contract and a few questions leave plenty of room. Two hours of back and forth fill the tray: it slows down, or loses track of what it read earlier.

Expect a wait, too. The model reads everything you gave it before it writes a word: one creator measured 41 seconds for a prompt of about 30,000 tokens (a token is about three quarters of a word), on a 16GB graphics card. A laptop is slower still. When Micro Center tested OpenAI's open model on a 16GB Mac mini, it came last, at about 13 to 18 tokens a second.

The rule. On a 16GB laptop, pick the 5 gigabyte class. For one document, slower is fine.

03 Laptop or cloud: sort the job

Is it as good as ChatGPT? Not overall: a model that fits on a laptop knows less and reasons less than the biggest cloud models. So judge it one job at a time.

On the laptop In the cloud

Summarize a client contract and list the dates and penalties Anything public

Turn notes from a sensitive call into a recap Anything that needs the web, like researching a competitor

Draft a reply to an upset customer, account details included Long, multi-step work, like an agent working through your inbox

Pull the key numbers out of a financial statement Anything that needs the strongest model there is

The pattern on the left: one document, one clear job, a few minutes, on a file that should never leave the building. Boring bulk work counts too.

Files follow the tray too. In LM Studio a short Word file, PDF or text file goes into the chat whole. A very long one gets searched for the relevant bits instead, which LM Studio's docs say sometimes needs tuning. One contract, yes. Your whole shared drive, no.

04 The 20-minute first test

Install one app from its official site. LM Studio from lmstudio.ai, or Ollama from ollama.com.

Download one model around 5 gigabytes, like Qwen3. In Ollama, the library page for Qwen3 shows the command: "ollama run qwen3:8b", the 5.2GB version.

Turn off your Wi-Fi. It still works. The file goes nowhere.

Give it one real document you'd never paste into ChatGPT. In LM Studio, attach the file to the chat. Keep it short: one contract, one set of notes.

Ask for the summary you'd actually use. One try tells you whether it's good enough for that job.

First test prompt

Summarize this [CONTRACT / CALL NOTES / STATEMENT] in [NUMBER] short bullet points.
Then list every date, every penalty and every key number it mentions.
Only use what is in the document.

05 What local doesn't fix

Local isn't automatically safe. You're still installing an app, downloading files from the internet, and sometimes adding plug-ins that can reach your folders. Treat a model like any other software.

Get the app from its official site. lmstudio.ai or ollama.com.

Get models from the app's own library. A model is a file you download from the internet.

Don't open it up to your network unless you know why. Local should mean this one machine.

Check your client contract before you promise anything. "Your files never left our office" is a strong promise. Make sure the contract agrees.

And no, I wouldn't cancel anything. Claude Pro is $20 a month, and the cloud is still where the hard thinking happens. The question just changes from "which AI should I pay for?" to "where should this file go?"

Sources

Ollama: Privacy Policy Ollama library: qwen3 LM Studio: System Requirements LM Studio: Chat with Documents Google Developers Blog: Gemma 3 QAT models AppleInsider: MacBook Air doubles base memory to 16GB Micro Center: Can your PC run OpenAI's GPT-OSS models? Kai: Is 16GB All You Need for Serious Local LLM? Malwarebytes: OpenAI forced to preserve ChatGPT chats Engadget: OpenAI no longer has to preserve all of its ChatGPT data Claude: Plans and pricing

Mathew's LibraryLibrary card

Free access

Keep reading free with your library card

One email opens this guide and all 53 resources in Mathew's library. No account, no password.