For Pistilli, opting for local models as opposed to online chatbots has implications beyond privacy. “Technology means power,” she says. “And so who[ever] owns the technology also owns the power.” States, organizations, and even individuals might be motivated to disrupt the concentration of AI power in the hands of just a few companies by running their own local models.
Breaking away from the big AI companies also means having more control over your LLM experience. Online LLMs are constantly shifting under users’ feet: Back in April, ChatGPT suddenly started sucking up to users far more than it had previously, and just last week Grok started calling itself MechaHitler on X.
Providers tweak their models with little warning, and while those tweaks might sometimes improve model performance, they can also cause undesirable behaviors. Local LLMs may have their quirks, but at least they are consistent. The only person who can change your local model is you.
Of course, any model that can fit on a personal computer is going to be less powerful than the premier online offerings from the major AI companies. But there’s a benefit to working with weaker models—they can inoculate you against the more pernicious limitations of their larger peers. Small models may, for example, hallucinate more frequently and more obviously than Claude, GPT, and Gemini, and seeing those hallucinations can help you build up an awareness of how and when the larger models might also lie.
“Running local models is actually a really good exercise for developing that broader intuition for what these things can do,” Willison says.
How to get started
Local LLMs aren’t just for proficient coders. If you’re comfortable using your computer’s command-line interface, which allows you to browse files and run apps using text prompts, Ollama is a great option. Once you’ve installed the software, you can download and run any of the hundreds of models they offer with a single command.
If you don’t want to touch anything that even looks like code, you might opt for LM Studio, a user-friendly app that takes a lot of the guesswork out of running local LLMs. You can browse models from Hugging Face from right within the app, which provides plenty of information to help you make the right choice. Some popular and widely used models are tagged as “Staff Picks,” and every model is labeled according to whether it can be run entirely on your machine’s speedy GPU, needs to be shared between your GPU and slower CPU, or is too big to fit onto your device at all. Once you’ve chosen a model, you can download it, load it up, and start interacting with it using the app’s chat interface.
As you experiment with different models, you’ll start to get a feel for what your machine can handle. According to Willison, every billion model parameters require about one GB of RAM to run, and I found that approximation to be accurate: My own 16 GB laptop managed to run Alibaba’s Qwen3 14B as long as I quit almost every other app. If you run into issues with speed or usability, you can always go smaller—I got reasonable responses from Qwen3 8B as well.