Running an AI at home, without subscription or cloud: the 5-step guide

Your AI, your machine, your data. Open AI models now run on a simple laptop, offline and without paying a subscription. For a crypto reader, the argument is obvious: nothing you type leaves your disk, at a time when AI agents are already touching wallets. Here's how to do it.

Step 1: check that your machine holds up

It all depends on memory. For small models, count on at least 8 GB of video memory on a graphics card, or an Apple Silicon Mac whose unified memory plays the same role, according to the guide published on November 1 by Decrypt. Serious uses require 24 GB of video memory or more.

LM Studio sets its own thresholds. On a Mac, you need an Apple Silicon processor, 16 GB of memory and macOS 14 at least, even if 8 GB Macs run the small models. On Windows, 16 GB of RAM, 4 GB of dedicated video memory and an AVX2 compatible processor. Intel Macs are not supported.

Step 2: Choose between Ollama and LM Studio

Two free tools dominate. Ollama is used on the command line and is mainly aimed at developers. LM Studio offers a classic graphical interface, with chat window, and works completely offline once the model is downloaded.

If you have never opened a terminal, take LM Studio. If you want to plug the model into your own scripts, Ollama exposes a local interface at localhost:11434.

Step 3: download and install a first template

With Ollama, one command is enough to install the tool, then two others to retrieve and launch a model: ollama pull gemma4:e2b, then ollama run gemma4:e2b. The /bye command closes the session. Beware of a trap, models suffixed “cloud” run on Ollama servers and not on your machine.

With LM Studio, everything goes through an integrated search engine. You choose a model, LM Studio indicates whether it fits in your memory, you click.

Step 4: Choose the Right Model for Your Machine

Google's Gemma 4 comes in several sizes, from 4.6 to 20 GB download; its 26B version requires 16 to 19 GB of memory. OpenAI's gpt-oss-20b runs with 16 GB, its big brother gpt-oss-120b requires up to 80. DeepSeek R1 is published under the MIT license, one of the most permissive.

The golden rule: the model must fit entirely in video memory. As soon as it overflows into traditional RAM, the speed plummets.

Step 5: Use it for what really matters

The uses that justify the effort are those where confidentiality takes precedence. Analyze a transaction statement or a tax table, summarize a contract, write a sensitive document. One absolute rule does not change with local AI: a wallet recovery phrase is never typed anywhere, neither online nor offline.