LM Studio

What it is

LM Studio is a desktop application for running open-weight language models on your own computer. It combines a model browser (searching and downloading from Hugging Face), a chat interface and a local server, so you can try a model in a window and then call the same model from your own code. It runs on Apple Silicon Macs, x64 and ARM64 Windows PCs and x64 Linux.

Under the hood it uses llama.cpp to run GGUF models on all three platforms, and Apple’s MLX on Apple Silicon. You can load models with different context lengths and GPU offload settings, connect MCP servers so local models can use tools, and chat with documents offline.

There are a few related pieces:

• lms, a command-line tool that ships with the app, for downloading, loading and serving models from the terminal. It is MIT licensed.
• llmster, a headless version of LM Studio for servers and CI, with no desktop app required.
• SDKs for Python (lmstudio-python) and TypeScript (lmstudio-js).
• LM Studio Bionic, a separate, newer agent app from the same company for coding, documents and research with open models. It can use local models or paid hosted open models. This page covers the original LM Studio app, which remains available alongside it.

The app itself is proprietary freeware made by Element Labs. Since July 2025 it has been free for use at work as well as at home, with no separate commercial licence needed. The terms mention optional paid features, and the paid plans on the pricing page relate to Bionic’s hosted inference rather than to running models locally.

Where it fits in AI-driven worlds

LM Studio covers much the same ground as Ollama, with a stronger focus on the graphical interface:

• Comparing models for a character. Before wiring a model into a game, you can load several small models in turn, give each the same persona prompt and see which voice and response time suits a shopkeeper or narrator.
• Offline prototyping of runtime AI. The local server lets a Godot project, a browser game or a Python simulation call a local model with no per-call cost and no data leaving the machine.
• Many-agent simulations. When hundreds of simulated agents each need to make decisions, a local model avoids a growing API bill.
• Local coding assistance. LM Studio documents integrations with Codex and Claude Code, so you can point a coding agent at a local model for private or offline work, accepting lower quality than hosted models.

Strengths

• Approachable interface. Searching, downloading and testing models needs no terminal, which helps designers who do not want to manage command-line tools.
• Hardware guidance. The model browser shows which builds should fit your machine, and load settings are exposed rather than hidden.
• Familiar API. OpenAI-compatible endpoints (chat completions, responses, embeddings) plus its own REST API, so existing client libraries work with a base URL change.
• MLX on Macs. Native MLX support often runs well on Apple Silicon.
• Free for work. No licence form or fee for commercial use of the app.

Limits

• Not open source. The app is proprietary, unlike Ollama or llama.cpp, so you cannot audit or modify it, and its terms can change.
• Hardware bound. Speed and quality depend on your RAM and GPU memory. Models that fit on a laptop are weaker than frontier hosted models at long coding tasks and complex reasoning.
• Not for shipping to players. It is a tool for developers. Running models on players’ machines needs a different approach, such as embedding a runtime directly.
• Model licences vary. LM Studio’s terms cover the app only; each model you download has its own licence, which you should check before commercial use.
• Product direction is shifting. With Bionic now the company’s headline product, expect the line between the two apps to keep changing.

Getting started

Download LM Studio for your system from lmstudio.ai/download, open it once, then search for and download a small model. To serve it to your own code, start the server from the Developer tab or from the terminal:

lms get openai/gpt-oss-20b
lms load openai/gpt-oss-20b
lms server start

The server listens on port 1234 by default. From Python, using the OpenAI client library:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:1234/v1", api_key="lm-studio")
reply = client.chat.completions.create(
    model="openai/gpt-oss-20b",
    messages=[{"role": "user", "content": "Describe a market town in two sentences."}],
)
print(reply.choices[0].message.content)

Choose a smaller model if your machine has limited memory.

Verdict

LM Studio is the friendliest way to explore open models locally, and its OpenAI-compatible server makes it easy to test runtime AI ideas without a bill. It is free for work but not open source, so Ollama may suit you better if that matters. For comparing models and prototyping offline characters, it is a good choice.