A language model you run yourself is two things on your own disk: a program to load the file, and the file. Nothing you type goes to a cloud where another company could inspect or analyse it, the model answers with the internet disconnected, and there is no subscription to pay and no request limit to watch. On Windows and macOS the shortest route for a beginner is LM Studio Bionic, which is free. The part that decides whether any of this works is memory.
What you give up is capability. Plenty of free models are available to download, including ones from Meta and Google, and they are generally less advanced and slower than the models inside paid apps — adequate for ordinary tasks, and you can keep several around and pick a different one per job. The other cost is maintenance: you install the updates yourself, and some of the polish of a finished ChatGPT app disappears. In exchange you get a personal system with far more control over your data, and getting started is genuinely simple.
Source: wired.com
Local models run on Windows, macOS and Linux. Among AI hobbyists macOS is the common choice, because Apple's hardware and software are more unified and Apple Silicon puts the CPU, the GPU and the memory together — an architecture that suits model work.
On any platform, more RAM is better, and the numbers matter more than the operating system does:
8GB — the stated minimum, with a restricted choice of models and slow output 16GB — the more workable figure 32GB or more — for the largest and fastest models More than 8GB of video memory — where a dedicated GPU starts to help noticeably
A dedicated graphics card with a lot of video memory gives the best results, because that memory is optimised for the kind of work models do. On Windows a dedicated Nvidia card is particularly useful: GPUs suit AI workloads better than ordinary CPUs, and their own memory gives the model extra room to operate.
There is no formal minimum requirement for running a local language model. But a large amount of RAM and a separate graphics card visibly widen what you can choose from. Beyond the hardware, you need two things: a program to run the model — the shell around it — and the model file itself. There are several options for each.
LM Studio Bionic is the usual recommendation for beginners on Windows and macOS. The other popular choices are vLLM, Llama.cpp, Ollama and GPT4All; they also run on multiple operating systems but take a more technical approach. Once you have picked an app, you pick a model. The app will suggest a few, and more are available in online catalogues — the best known being Hugging Face, which holds more than 3 million models.
Source: wired.com
The configurations are many; here is the basic Windows one. Download the installer from the app's site and launch LM Studio Bionic. Click Create project and give your first project a name.
An empty chat window opens. Click Select model in the prompt field, then Get local models, and the app opens its main model catalogue. Every entry comes with its size, its popularity and its characteristics. For a first run, the app team's own recommendations are a reasonable place to start. Smaller models are usually less capable, but they download faster and take up less space.
The chat window works much like any other assistant app. The prompt field at the bottom lets you switch between the models you have installed. If a model handles images and files, a "+" icon appears to the left of that field for attaching them.
The navigation panel on the left switches between projects. Bottom left there is a Settings link, which opens LM Studio Bionic's settings — among them how deleted chats are handled and which interface elements are shown.
Source: wired.com
Inside settings, the Library section manages the models you already have, and Browse searches for new ones. If you want to work with images and documents, choose multimodal models; technically they extend a language model's capabilities, though they are usually called language models too.
Finally, the sidebar on the right, shown or hidden with the button in the top right corner, holds additional tools for managing files across projects and for granting the program access to your computer's file system if you need that.
Read the memory table again, because it is the whole piece. An 8GB minimum that comes with a restricted model list and slow generation is not really a minimum; it is a demonstration. The practical entry point is 16GB, and the number that changes what you can run is 32GB. That reframes the pitch: this is not privacy versus convenience, it is privacy versus capability, and the trade is stated openly — less advanced, slower. Anyone switching to a local model to escape a subscription pays the same bill in tokens per second and in answer quality, or pays it up front in memory.
The guide is quiet about the two things a beginner hits first. How large are these files, and how much disk do they consume? Smaller models "take up less space" is the only answer offered, and space is what runs out. And which model, specifically? Meta and Google are named as sources, and Hugging Face is named as the catalogue, but a catalogue of more than 3 million files is not guidance — it is the precise moment at which people give up and reopen ChatGPT.
The installation takes a few clicks and is the easy part. Choosing among three million files is the part nobody has made easy, and it decides whether the model sitting on your machine is useful or merely private.