LM Studio

LM Studio is a desktop application for macOS, Windows and Linux that lets people download and run large language models (LLMs) on their own computer instead of sending requests to a remote service, and experiment with them locally.^[1]

What the app contains

The program bundles a set of tools in one window: a chat interface for talking to models, a way to search for and download models from Hugging Face, a local server that listens on endpoints shaped like OpenAI's, and systems for managing the models stored on the machine and their configurations.^[1] The point of putting all of this in one application is that a user does not have to assemble a separate inference engine, model downloader and server by hand.

How models run

On all three supported platforms LM Studio executes models with llama.cpp, using the GGUF model format. On Macs built around Apple Silicon it can additionally use Apple's MLX. These execution backends are managed inside the application as "LM Runtimes".^[1]

llama.cpp is an open-source inference engine for running language models on ordinary hardware, and GGUF is a file format for storing such models, usually in compressed (quantized) form so they fit in the memory a personal computer actually has. Because inference happens on the user's own machine, the model runs on the local processor and graphics hardware rather than in a data centre, and no request has to leave the computer.

Offline documents and local interfaces

Running everything locally means the app can be used to chat with the user's own documents entirely offline, a retrieval-augmented-generation (RAG) workflow: relevant passages are retrieved from the documents and passed to the model together with the question, so the answer can be grounded in that material rather than only in what the model memorised during training.^[1][2]

LM Studio also exposes the loaded models to other programs. It provides an OpenAI Compatibility API and a beta REST API for local models, so software written against a hosted chat service can be pointed at the machine instead.^[1][2] The documentation shows the REST API being called at `http://localhost:1234/api/v0/chat/completions`, an address that makes clear the server is listening on the user's own computer.^[1]

Releases and licensing

The desktop app is proprietary and closed-source.^[1] Its documentation changelog lists named releases; one example is LM Studio 0.3.9, released on 30 January 2025, which added an idle time-to-live (TTL) and an auto-evict setting for models loaded through API requests, so that a model could be unloaded automatically after sitting unused for a set period.^[1]

Where it sits

LM Studio belongs to the group of tools that bring model inference down to consumer hardware. Its distinguishing combination is a graphical interface, built-in model search and management, and OpenAI-compatible local endpoints, which let existing tools treat a model running on a laptop as though it were a cloud service. The trade-off of the local approach is that what a user can run is bounded by the memory and computing power of the machine in front of them, and the application itself is not open source, so its internals cannot be inspected or modified the way open-source inference engines can.^[1]