LM Studio is a desktop app that downloads open-weight language models and runs them on your own machine. It exposes a local server that speaks the OpenAI API shape, so anything you wrote against a hosted model points at localhost with a one-line change.

How I use it

Two jobs. The first is bulk classification: tag a few thousand ad headlines by angle, sort scraped pages into buckets, decide which of two hundred transcripts mention a product. These tasks are easy per item and expensive in volume, and a local model does them for the price of electricity. The second job is first drafts. A local model writes a rough outline or a rough translation, and the hosted model gets the polished pass. My rule is simple: before I pay for a bulk API call, I check whether a local model can do it well enough, and often it can.

What surprised me is how usable a mid-size model is on a Mac with a lot of unified memory. Responses come back fast enough for a loop over a file, and the machine stays usable. What annoys me is the gap at the top: for anything subtle, a nuance in tone, a tricky judgment call, the local model is a tier below and you feel it. And the catalogue needs curation. Searching for a model returns twenty quantised variants of the same weights, and picking the right one takes a few tries.

The CLI is the part people miss. You can load and unload models from a script and keep the GUI closed, which is what makes it a real tool rather than a demo.

Who it is for

Anyone with a recent Mac or a decent GPU who is curious what local inference can do. It costs nothing and takes ten minutes to find out. Skip it if your machine has 16 GB or less, or if everything you do needs the best model available; then the hosted APIs are the better deal and the app will only disappoint you.

Related tools