Power & Pantry

Disclosure: Power and Pantry is reader-supported. If you buy through links on this page, we may earn a commission from Amazon or another program at no extra cost to you. See the About page for how this site is funded.

How to Run Your Own AI When the Internet Is Down (2026)

The short answer

Yes, if you set it up before you need it. A small free program, Ollama or LM Studio, runs an AI model on your own laptop or desktop. You download the model once, while you still have a connection; the examples on this page run from 1.3 to 4.7 gigabytes. After that download, the model answers with the Wi-Fi off, because the work happens on your machine and nowhere else. ChatGPT and the other big-name chatbots cannot do this. They run on their companies' servers, there is no model file to download, and when the connection goes, they go with it. A local model is smaller and less capable than the cloud services, it knows nothing that happened after its training, and it can be wrong with complete confidence. What it offers in an outage is a private helper that still works: summarizing, drafting, rewriting, translating, and explaining text you already have. The setup ends with the proof, and the proof is simple. Turn the Wi-Fi off and ask it something.

Why the AI you know stops at the connection

ChatGPT, Claude, and Gemini are services, not programs you own a copy of. The model lives on the company's servers, your words travel to it, and the answer travels back. That design is why they are as capable as they are, and it is also why no setting, subscription, or app version makes them work offline. There is nothing on your computer to run. The offline kind of AI is a different thing: an open-weight model, a file you download and keep, run by a program on your own machine. Open-weight means the model's maker publishes the file for others to download and run, under that maker's license. Meta's Llama models are the examples used on this page. The trade is plain. A model small enough to fit on a household laptop is much smaller than the server models, and it answers like it: useful for everyday language work, weaker on hard reasoning, and with no knowledge of anything after its training date.

The readiness point follows from the file. Everything about this setup happens in advance, on a working connection, on an ordinary day. The programs are downloads of their own, the models are gigabytes each, and none of it can be fetched during the outage it is meant for. Treat this page the way this site treats a water container or a charged power station. Fill it before you need it.

The two tools, and how they differ

Two programs cover nearly every plain-language guide to local AI, and both are documented well enough to set up from their makers' own instructions. Ollama is the terminal-first option: you install it, type one command naming a model, and it downloads the model and opens a chat. It also ships a desktop app on Mac and Windows. Ollama is free and open source, published under the MIT license, and it runs on macOS 14 Sonoma or newer, Windows 10 22H2 or newer, and Linux. On a Mac with Apple silicon it uses both the processor and the graphics hardware; on an Intel Mac it runs on the processor alone, which is slower. The Windows install does not need administrator rights. LM Studio is the graphical option: a normal app window with a built-in model search and downloader, a chat pane, and settings you can see. It is free of charge; the company announced in July 2025 that it is free to use both at home and at work, with no form or approval needed. It is not open source, which is the honest difference between the two tools: Ollama's code is public, LM Studio's is not. LM Studio's own system requirements are narrower on the Mac, where it supports Apple silicon machines only, M1 through M4, on macOS 14.0 or newer, and Intel Macs are not supported at all. On Windows it needs a 64-bit machine whose processor supports AVX2, or an ARM machine such as a Snapdragon X Elite laptop. On Linux it is distributed as an AppImage and requires Ubuntu 20.04 or newer.

ToolCostOpen sourceHow you use itRuns on
OllamaFreeYes, MIT licenseOne-line commands, plus a desktop appmacOS 14+, Windows 10 22H2+, Linux
LM StudioFree of charge, home and workNoA graphical appmacOS 14+ (Apple silicon only), Windows x64 or ARM64, Linux

Both tools need the internet exactly once in this setup, to fetch the program and the model. Ollama pulls models from its own library. LM Studio's downloader searches Hugging Face, the public model repository, from inside the app. Both then run the model on your machine: LM Studio's documentation describes running models through the llama.cpp and MLX runtimes on the computer itself, and Ollama's FAQ states its position in one sentence worth quoting in substance. Ollama runs locally, and the company does not see your prompts or data when you run locally.

Size the model to your RAM, not to your ambitions

A model has to fit in memory to run, and the memory that decides the fit is your computer's RAM, shared with everything else the machine is doing. Model names carry their size class in billions of parameters: a 3B model is a small one, an 8B model is a mid-size one, and the number is the first thing to match against your machine. Ollama's own README gives the ladder in plain terms: have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models. Read the word available carefully. Your operating system and open programs are already using some of your RAM, so this page plans one class under the ladder's minimums, and LM Studio's documentation points the same way: it recommends 16 GB or more, and says an 8 GB Mac should stick to smaller models in the 1B to 3B range with modest context sizes.

RAM in your computerModel class to plan onExample from Ollama's libraryDownload size, as listed
8 GBSmall models, 1B to 3BLlama 3.2, 1B or 3B1.3 GB or 2.0 GB
16 GBThe 7B to 8B classLlama 3.1, 8B4.7 GB
32 GBThe 13B class, and up to the low 30sPer Ollama's ladder, 33B models ask for 32 GBVaries by model; the library lists each one

The download sizes above are Ollama's listed sizes for the default builds, checked in November 2026. Two things about them deserve a sentence each. First, the default builds are compressed, using a technique called quantization that shrinks the model file in exchange for some quality; LM Studio's documentation advises choosing a 4-bit option or higher, and 4-bit is what Ollama's default tags ship. An uncompressed build of the same model can be three times the size, which is why the small numbers in the table are real and not a trick. Second, listed sizes move slightly when a model build is updated, so the figure that counts is the one shown when you download. A graphics card, or the graphics half of an Apple silicon chip, makes answers faster, and Ollama documents what happens without enough of that faster memory: the model falls back to ordinary RAM and answers slow down. Speed is a comfort here, not a requirement. A slow answer during an outage is still an answer.

One more size limit is invisible until it bites. A model can only hold so much text in mind at once, a window measured in tokens, which are word pieces. Ollama's documented default context window is 4,096 tokens, roughly a few thousand words of conversation and pasted text. Long documents fill it fast, and what falls out of the window is forgotten mid-conversation. For outage work this argues for short, specific questions and pasted excerpts over whole manuals at once. Feeding a local AI your own documents in a managed way is a separate job, and a later page in this corner takes it up.

Setup path one: Ollama, as documented

  1. While you have a connection, download Ollama from its official site for your system, and install it. On Linux, the documented route is the install script on that same download page, which sets Ollama up as a system service.
  2. Open a terminal and run one command naming the model that fits your RAM from the table above. For most laptops that is ollama run llama3.2, which downloads the 3B model, about 2.0 GB, and opens a chat when the download finishes. On an 8 GB machine, ollama run llama3.2:1b fetches the smaller 1B model instead.
  3. Type a question and confirm you get an answer. The model file now lives on your machine, in the .ollama folder in your home directory, and ollama list will show what you have downloaded.
  4. Optional, and recommended for this use: switch on Ollama's documented local-only mode, which disables its cloud features entirely. The FAQ describes two ways, setting disable_ollama_cloud in the server configuration file, or the OLLAMA_NO_CLOUD environment variable, and notes that the program's log then prints that cloud is disabled. This matters because Ollama also offers cloud models, whose names carry a -cloud tag and which run on Ollama's servers. Those are the opposite of this page's purpose, and local-only mode removes them from the picture.
  5. Finish with the verification step below, Wi-Fi off, before you call the setup done.

Setup path two: LM Studio, as documented

  1. Check your machine against LM Studio's stated requirements first, because they exclude real machines: no Intel Macs, macOS 14.0 or newer, 16 GB of RAM recommended, and on Windows a processor with AVX2 support or an ARM chip.
  2. Download LM Studio from its official site and install it like any other app.
  3. Use the built-in downloader to fetch a model. The documentation describes searching by keyword from inside the app, with the model files coming from Hugging Face, and advises picking a 4-bit build or higher for machines that can manage it. Pick a model in the size class your RAM allows, using the same table above; the model families overlap with Ollama's library, Llama included.
  4. Load the model in the chat pane and ask it something, to confirm the download is complete and the machine runs it.
  5. Finish with the same verification step below, Wi-Fi off.

A note on accounts and licenses, since both tools are free and that invites the question of what the catch is. For local use, neither setup asks you to create an account: Ollama's sign-in exists for its cloud models, which this page does not use, and LM Studio's downloader works inside the app. The models are a separate matter from the tools. Each model carries its own license from its maker, printed with the model in the library, and Meta's Llama models are distributed under Meta's own community license rather than an open source one. For a household readiness setup the practical terms are simple, but anyone planning to use a local model in a business should read the model's license first, because the tool being free does not make every model free for every use.

The verification step: turn the Wi-Fi off

Every setup guide for local AI says the result works offline. This one asks you to prove it on your own machine, the same day you set it up, while proving it costs nothing. With the model downloaded and a chat open, turn off the Wi-Fi and unplug any network cable, or switch on airplane mode. Then ask a fresh question, something you have not asked in this session. If an answer comes back, the model on your machine produced it. There is no other machine it could have come from. Ollama adds two documented confirmations for readers who want them: the command ollama ps lists the running model and shows whether it is loaded on the graphics hardware, the processor, or split between them, and Ollama serves its models only to the machine itself by default, at the local address 127.0.0.1, port 11434.

Be clear about what the check does not prove. It proves the chat path is local. It does not prove the program never touches the internet when you are connected; Ollama's FAQ describes its desktop apps downloading updates automatically, with a setting to only check instead, and update checks are a normal thing for software to do on a working connection. The privacy claim on this page is the narrower, documented one: with a local model, your prompts and the model's work stay on your machine, and in Ollama's local-only mode the cloud features are off entirely. What you type into a local model during an outage goes nowhere, because there is nowhere for it to go.

What it can do during an outage, and what it cannot

A small local model earns its disk space on language work you feed it. It can summarize a long document you paste in pieces, rewrite a note into a clearer message, draft a letter or a list from your bullet points, explain a passage in plainer words, and translate between languages; the Llama 3.2 models in the table officially support eight, including English and Spanish. It can talk through a problem with you at 2 a.m. when there is no one to call and nothing to load. All of that works at the speed of your machine, for as long as your machine has power, with nobody else involved.

The limits are just as concrete. It has no live information of any kind: no weather, no road conditions, no news, no idea an outage is even happening. Its knowledge stops at its training, and it fills gaps by guessing in the same confident tone it uses for facts, so any specific claim it makes, a dosage, a wiring detail, a rule, has to be checked against a real reference before you act on it. It cannot browse, send, call, or control anything. For word of what is actually happening, the tools are the ones in this site's off-grid communication explainer and its emergency radio comparison: a local AI explains what you already have, and a radio tells you what you do not. The pairing that makes a local model genuinely useful in a long outage is a downloaded reference library next to it, frozen facts it can help you find and understand. A later page in this corner covers building that library.

Power: count it like any other load

A computer running a model works harder than a computer idling, and on a laptop that means the battery drains faster, by an amount that depends on the machine and the model. This page prints no watt figure for local AI, because neither maker publishes one and the honest number is measured, not looked up; the next page in this corner takes up exactly that question, using the figures the hardware makers do publish. Until then, plan with the figure you already own: the wattage printed on your laptop's charger is the most the machine can draw, and entering that number in this site's generator wattage calculator gives a cautious runtime estimate for AI work alongside your other loads. The planning rule is the one every power page here uses: a station delivers about 85 percent of its rated capacity, so a 1 kilowatt-hour station offers about 850 usable watt-hours. A laptop is a modest load next to a refrigerator, and an hour of questions is cheap. Leaving a model loaded and a laptop awake around the clock is a different decision, and the calculator will show you its cost in hours. The stations themselves are compared in this site's portable power station comparison.

Matched offers are coming. This site is new, and the affiliate links for this corner are still being set up. When they are live, this box will hold current offers for the hardware this corner's pages name, such as external drives for storing models and offline libraries, with prices re-checked at that time. The disclosure at the top of this page explains how the site earns.

What we could not verify

Questions

Can ChatGPT work without internet?

No. ChatGPT runs on OpenAI's servers, and there is no downloadable version of the model for a home computer. The app and the website are windows onto a service; with no connection, there is nothing behind the window. The same holds for Claude, Gemini, and the other hosted chatbots. Working without internet means running a different, smaller model locally, which is what this page sets up.

Does AI need the internet?

The hosted kind does, for every message. A local model needs the internet once, to download the program and the model file, and after that it does not. That single difference is the whole subject of this page, and the Wi-Fi-off check above is how you confirm which kind you are talking to.

How much does this cost?

Nothing in money. Ollama is free and open source, LM Studio is free of charge for home and work use, and the example models are free to download under their makers' licenses. The real costs are disk space, 1.3 to 4.7 GB for the models in the table, the electricity to run the machine, and twenty minutes of setup on a day when nothing is wrong. Ollama sells cloud plans; this setup does not use them and does not need an account.

Can I run a local AI without a graphics card?

Yes. Both tools run on the processor alone, which is how every Intel Mac runs Ollama, and the answers come slower. Stay with the small models in the table, expect short answers to arrive at reading speed rather than instantly, and the setup remains useful. A graphics card or an Apple silicon chip is a speed upgrade, not an entry requirement.

Will it know what is happening during an emergency?

No. It knows what was in its training, which ended before your emergency started, and it has no way to receive updates with the connection down. Ask it to explain, summarize, draft, and translate. Get conditions, warnings, and instructions from official sources and a radio, and check anything the model states as fact against a reference you trust before acting on it.

Is what I type into it private?

With a local model, yes, in the specific sense that matters here: the conversation is processed on your machine and is not sent to the tool's maker. Ollama's FAQ says so directly for local running, and its local-only mode turns the cloud features off altogether. LM Studio likewise runs the model on your computer through local runtimes. The exception to keep straight is any model whose name marks it as a cloud model; those run on the maker's servers by design, and this page's setup avoids them.

What about AI on my phone, with no internet?

Phones have their own offline AI apps, which run smaller models under tighter battery and heat limits than a laptop allows. That is a separate comparison, and a later page in this corner covers it on its own terms. If your phone is your only computer, read this page for the shape of the idea and that page for the picks.

A note on responsibility

A local AI is a language tool, not an authority. It will produce fluent, confident text that is sometimes wrong, and an outage is the worst time to discover which kind you got. Do not take medical, electrical, structural, or safety instructions from a model, local or otherwise. Use it to work with information from sources that carry responsibility for theirs: official guidance, your downloaded references, and the manuals for your own equipment. This page's data basis is stated at the top and stands: official documentation and the makers' own library figures, checked in November 2026, with the gaps named where a reader would otherwise assume a test or a number exists.