LM Studio: what it is, what it is for, and when to use something else
A desktop GUI for discovering, running and chatting with local models.
At a glance
| Capability | LM Studio |
|---|---|
| Licence | Proprietary (free for personal use) |
| Implementation | Electron desktop app |
| Runs on | CPU or GPU |
| Continuous batching | No |
| Paged attention | No |
| Tensor parallelism | No |
| Quantisation | GGUF, MLX (Apple Silicon) |
| OpenAI-compatible API | Yes |
LM Studio is a desktop application for running open models locally with a graphical interface. It bundles model discovery from Hugging Face, quantisation selection, a chat UI, and a local server mode that exposes an OpenAI-compatible endpoint. On Apple Silicon it can use MLX in addition to GGUF, which materially changes throughput on Mac hardware.
Its practical advantage over command-line runners is evaluation speed. Comparing a 4-bit and an 8-bit build of the same model, or two models at the same quantisation, is a few clicks rather than a scripted benchmark, and the interface surfaces memory usage as you load. For deciding what will actually fit on a given machine, this is faster than reasoning about it on paper.
It is not a deployment target. The licence is proprietary, the distribution model is a desktop installer, and there is no clustering story. Treat it as a workstation tool for choosing a model, then serve that model with something built for serving.
Where it stops being the right tool
A desktop application, not a server product. It is not something you deploy, and its licence is not open source.
Frequently asked
- What is LM Studio used for?
- Non-terminal users, model evaluation, and anyone who wants to compare quantisations of the same model interactively before committing to one.
- What are the limitations of LM Studio?
- A desktop application, not a server product. It is not something you deploy, and its licence is not open source.
- Does LM Studio need a GPU?
- No. LM Studio runs on CPU, and offloads to a GPU when one is available. CPU execution is much slower but works for evaluation and small models.
Compare with
- Ollama — The simplest way to run open models on your own machine.
- vLLM — High-throughput inference server built around PagedAttention.
- llama.cpp — Portable C/C++ inference that runs open models almost anywhere.
- SGLang — Inference server optimised for structured generation and prefix reuse.
Project home: https://lmstudio.ai