Guide
How to use an AI model on your computer, without internet or fees
By the end, you'll know what a "local" AI model is, whether your computer can handle one, and how to install it with free programs so you can chat with it offline and without paying for a subscription.
Translated with AI from the Spanish original. Nobody has reviewed this translation; if anything seems off, the Spanish version is the reference.
What you'll get
When you use ChatGPT, Gemini or Claude, your question travels to those companies' computers, gets processed there, and the answer comes back to you. You need a connection, and the most powerful features usually cost money.
A local model works the other way round: the program and the AI are inside your computer. After an initial download, you can turn off the wifi and keep using it. What you type never leaves your machine, and there's no monthly fee.
You can ask it for the usual things: write an email, summarise a text you paste in, explain a concept, translate, or check spelling. Don't expect the same as you'd get from the most advanced paid assistants. This guide explains why.
Before you start: what "open" means
To run an AI at home, the company that created it must have published its "weights". These are the billions of internal numbers the model learned during training, and they decide how it responds. If the weights are public, anyone can download them. This is called an open-weight model.
It's not quite the same as "open source". Many companies publish the weights but not the data they used to train the model, and some set conditions of use in their licence. If you're going to use it for a business, read that licence first.
There are several well-known families of open models: Llama (from Meta), Gemma (from Google), Qwen (from Alibaba), Mistral (from the French company of the same name) and DeepSeek (from the Chinese company of the same name). The available versions change over time, so this guide doesn't recommend any specific one.
What you need
A recent computer running Windows 10 or 11, macOS 14 or later, or Linux. If you have a Mac with an Intel processor, LM Studio won't work on it; Ollama will, although more slowly.
RAM is what matters most. The figure attached to a model, for example "8B", tells you how many billions of parameters it has. Parameters are those internal settings. The more it has, the more memory it needs.
As a rough guide, with 8 GB of RAM, small models of about 1 to 4 billion parameters run smoothly. With 16 GB, models of about 7 or 8 billion work well. A powerful graphics card or a Mac with an Apple chip makes it respond considerably faster.
You also need free disk space. A small model takes up between 1 and 5 GB. The largest models released openly have hundreds of billions of parameters and won't fit on a home computer: they need professional equipment.
Finally, you need internet just once, to download the program and the model.
Step by step
There are two widely used free programs for this: LM Studio and Ollama. LM Studio has a window like any other app and is the simplest option if you've never done anything like this. Ollama is very popular and lightweight, and has an app with a chat window on Windows and Mac; on Linux it's used from the terminal, the text window where you type commands. The following steps work for both, with small differences.
- Check how much RAM your computer has. On Windows, open Configuración > Sistema > Información. On a Mac, open the Apple menu > Acerca de este Mac and also check the chip: if it says Intel, use Ollama.
- Download LM Studio from lmstudio.ai/download or Ollama from ollama.com/download. Only use the official website, and install it like any other program. The LM Studio website also offers another app, LM Studio Bionic: that's not the one this guide describes.
- Open the model search. In LM Studio, it's the section for discovering models, with the magnifying glass icon. For Ollama, browse its catalogue at ollama.com/library.
- Choose a small model from one of the well-known families that suits your RAM. LM Studio usually warns you if a model is too big for your computer.
- Download it and wait for it to finish. It may take several minutes.
- Turn off the wifi and type a question in the chat. If it answers, the AI is running entirely on your computer.
- When you no longer use a model, delete it from within the program to free up disk space.
What to expect
The first answer may take a few seconds while the model loads into memory. After that, it will write at a speed that depends on your computer. On a modest laptop it may go at the pace of slow reading, and the fan will probably kick in.
Small models make more mistakes than the big paid assistants. They invent facts more easily, reason less well on long problems, and know less about uncommon topics. They usually work well for writing, rephrasing or summarising a text you give them. They're less reliable for questions about specific facts.
They also know nothing about what happened after their training and, offline, they can't search the internet. If you ask them about something recent, they may answer confidently and be wrong.
As of November 2026, Ollama and LM Studio Bionic also offer models that run on the company's servers, not on your computer. Those models are paid for separately, and your questions go out over the internet. In Ollama they're marked as "cloud". If you want privacy and no cost, check that the model is downloaded on your computer.
When it's worth it
It makes sense if you want to work with private texts, such as personal notes or documents you don't want to upload to any company. It also makes sense if you travel without a good connection, or if you're simply curious to see how an AI works on the inside.
If what you need is the best possible quality, up-to-date information or analysis of very long documents, a cloud assistant will still do it better. Many have a free version. You can use both kinds.
Watch out: Only download programs and models from official websites or from well-known accounts on Hugging Face: modified versions of open models are circulating with their safety filters removed. Also, the model can be wrong even when it sounds sure of itself, so check any important fact.
Sources
- LM Studio, descarga
- LM Studio, requisitos del sistema
- LM Studio, presentación de LM Studio Bionic
- Ollama, descarga
- Ollama, requisitos en macOS
- Ollama, catálogo de modelos
- Hugging Face, repositorio de modelos
- WIRED (octubre de 2026)