LM Studio
Why it matters
LM Studio is one of the two most common on-ramps to local AI, alongside Ollama. Where Ollama is command-line first and popular with developers, LM Studio is a graphical app aimed at people who want to see what's happening: which models fit in memory, how fast they run and what settings they're using.
It's also a practical developer tool. Its local server exposes an OpenAI-compatible API, so you can point existing apps, coding tools or scripts at a model running on your own machine instead of a cloud provider. That's useful for privacy, offline work, cost control and testing.
Deep Dive
LM Studio is a free desktop application for macOS, Windows and Linux that makes local AI feel like an ordinary app. You search for a model, pick a size that fits your machine, click download and start chatting, without touching a terminal. Under the hood it uses llama.cpp to run GGUF models on almost any hardware, and Apple's MLX on Apple Silicon Macs.
It's made by Element Labs. The app itself is closed source, but its developer SDKs and command-line tool are open source. Since mid-2025, it's also free to use at work, not just for personal use.
What You Get
- Model discovery: search Hugging Face from inside the app, with hints on which quantized versions will fit in your RAM or VRAM.
- Chat interface: conversations, system prompts, document attachments for simple RAG, and side-by-side settings.
- Local server: an OpenAI-compatible endpoint (by default on
localhost:1234) that other software can call. lmscommand-line tool: load and unload models, start the server and script it all.- SDKs: Python and TypeScript libraries for building apps on top of local models.
- Tool use and MCP: support for function calling and connecting MCP servers, so local models can use tools.
Choosing a Model
The most common question is which file to download. Model pages offer several quantizations of the same model: Q4 versions use about a quarter of the original memory and are the usual sweet spot; Q8 is closer to full quality but twice the size; Q2 and Q3 save more memory at a noticeable quality cost. As a rough guide, a 4-bit model needs a bit more than half a gigabyte per billion parameters, plus extra for the context.
Settings That Matter
- GPU offload: how many layers run on the GPU. Put everything on the GPU if it fits; partial offload works but is slower.
- Context length: longer contexts need more memory for the KV cache. Don't set it far higher than you need.
- Engine: on a Mac, try both the MLX and llama.cpp (GGUF) versions of a model. MLX is often faster.
LM Studio Versus Ollama
Both run the same kinds of models, and both offer local APIs. LM Studio is better if you want a visual app, easy model browsing and hands-on settings. Ollama is better if you want a lightweight background service, scripting and open-source code. Many people use both.