How to run an AI chatbot on your own computer
Large language models, the technology behind chatbots like ChatGPT and Gemini, do not have to run in the cloud. Wired published a guide explaining that they can be downloaded and run entirely on a person's own computer instead, walking through why someone would want to, what it takes, and a full worked example. The appeal is straightforward: nothing typed into the chatbot leaves the machine for a company to review, there is no monthly subscription or usage cap, and many capable models, including ones from Meta and Google, can be downloaded for free. The tradeoff is maintenance. A locally run model needs its updates handled by hand, and it will usually be less advanced and slower than the paid, cloud-hosted version of a similar tool, though the guide says the free local options are capable enough for everyday use.
There is no fixed minimum specification for a computer that can run a local LLM, but more memory and a discrete graphics card both help. 8 GB of RAM is called the bare minimum, workable but limiting on model size and speed; 16 GB is described as better; and 32 GB or more is said to be required for the biggest, fastest models available. On the graphics side, a dedicated GPU with more than 8 GB of VRAM is said to make a real difference, since that memory is built for the kind of work AI models do. The guide points Windows users toward a dedicated Nvidia GPU in particular, noting how closely Nvidia's hardware is tied to the AI boom, while it calls macOS the platform most AI enthusiasts favor, because Apple Silicon chips combine CPU, GPU and RAM into a single unit.
Running a model locally needs two pieces: an app to serve as the interface, and the model file itself. LM Studio Bionic is named as generally the best app choice for both Windows and macOS, and it is free. More technical, multi-platform alternatives include vLLM, Llama.cpp, Ollama and GPT4All. For the model itself, several online repositories offer choices; Hugging Face is called the best known, and the guide says it hosts more than 3 million models.
To show how the pieces fit together, the guide walks through installing LM Studio Bionic on Windows. After downloading and opening the app, a user clicks Create Project and names it, then, from the blank conversation window, clicks Choose a model in the prompt box and Get local models to reach the main model picker, where every entry is listed with its size and popularity and staff picks are flagged for newcomers; smaller models download faster and use less space but do less. From there, the chat interface works like any standard AI chatbot: the prompt box lets a user switch between installed models and, for models built to support it, attach images or files; a left-hand navigation pane switches between projects and opens Settings, where a Library section manages installed models and an Explore section finds new ones; and a right-hand sidebar, which can be shown or hidden, manages files across projects and can be given access to the computer's own file system. The walkthrough covers Windows only; the guide does not include an equivalent step-by-step for macOS or Linux.
Key facts
- Running an LLM locally trades some speed and capability for offline privacy and no subscription cost, since many models from companies including Meta and Google can be downloaded free.
- RAM needs scale with ambition: 8 GB is the bare minimum, 16 GB is better, and 32 GB or more is needed for the biggest, fastest local models; a GPU with more than 8 GB of VRAM helps further.
- macOS is the platform most AI enthusiasts prefer because Apple Silicon unifies CPU, GPU and RAM, while Windows users benefit most from pairing a dedicated Nvidia GPU.
- LM Studio Bionic is named the best free app choice for Windows and macOS; vLLM, Llama.cpp, Ollama and GPT4All are named as more technical, multi-platform alternatives.
- Hugging Face, called the best known model repository, is said to host more than 3 million downloadable models, though the guide does not recommend a specific one.
Why it matters
Large language models are reshaping coding, writing, summarizing and search, and that shift shows no sign of slowing. Being able to run the same underlying technology entirely offline, on hardware a person owns, matters to anyone who wants a private assistant that does not send what they type to a company's servers, and who wants to avoid a recurring subscription. Free models from major labs, including Meta and Google, make that possible without paying for access, even though they are not the fastest or most advanced options available.
Who it affects
Anyone currently paying for or relying on a cloud chatbot who wants more privacy or a lower ongoing cost, at the price of doing their own maintenance and updates. It also affects anyone shopping for hardware with local AI in mind: the guide steers Mac buyers toward Apple Silicon, since it unifies CPU, GPU and RAM, and steers Windows buyers toward pairing a dedicated Nvidia GPU with as much RAM as they can afford.
How to use it
Two things are needed: an interface app and a model. LM Studio Bionic is named generally the best choice for Windows and macOS and is free; vLLM, Llama.cpp, Ollama and GPT4All are named as more technical alternatives that run on multiple operating systems. Models come from online repositories, chiefly Hugging Face, said to host more than 3 million of them. As a worked example, the guide installs LM Studio Bionic on Windows: create a project, then in the prompt box use Choose a model and Get local models to pick one, noting its size and popularity; the resulting chat interface supports switching between installed models, attaching files or images to models built for it, and a Settings menu with a Library for installed models and an Explore section for finding new ones. The walkthrough covers Windows only, with no equivalent step-by-step given for macOS or Linux.
How solid is it
This is a practical how-to piece, not a study or a product announcement, and it does not tie its advice to any specific recent launch or news event. Its hardware guidance is explicitly framed as a rule of thumb rather than a measured benchmark: the guide states there is no fixed minimum spec, only that more RAM and a discrete GPU help, and none of its figures, 8 GB, 16 GB, 32 GB of RAM, or the more than 8 GB VRAM threshold, come with tokens-per-second or latency numbers attached.
Risks and caveats
The guide is upfront that free local models tend to lag their paid, cloud-hosted counterparts in capability and speed, and that running one yourself means taking on maintenance, including updates, that a hosted chatbot app would otherwise handle. It names LM Studio Bionic as explicitly free but does not state the pricing or licensing of vLLM, Llama.cpp, Ollama or GPT4All, and it does not recommend any specific model to download, leaving that choice to the reader.