What is LM Studio?
LM Studio is a desktop app from Element Labs for Mac, Windows and Linux that finds open models on Hugging Face, downloads them and runs them on your own hardware. It runs GGUF models through llama.cpp on every platform and, on Apple Silicon Macs, MLX models through Apple's MLX framework. You can chat with a model, attach documents and work fully offline once the files are on disk.
The same app doubles as a local server. It answers OpenAI-compatible requests (chat completions, responses, embeddings) and Anthropic-compatible /v1/messages, has its own REST API, and ships TypeScript and Python SDKs and the lms command-line tool. It has acted as an MCP host since version 0.3.17. Version 0.4.0 added llmster, a headless daemon for machines with no screen, and parallel requests; 0.4.28 is the version on the download page.
The app has been free for home and work use since 8 July 2025. It is not open source, though the lms CLI and both SDKs are MIT-licensed on GitHub. Element Labs also sells Bionic, a separate agent app with paid cloud models.
Is LM Studio worth it?
Sending every prompt to a cloud chatbot is fine until the prompt is a client contract, a patient letter or source code you are not allowed to share. Then the choice is usually between not using AI at all and asking legal.
LM Studio runs the model on your own computer instead. You search for an open model inside the app, download it and chat with it. Once the file is on disk, nothing you type leaves the machine, according to the privacy policy, which also says the app carries no telemetry. It handles both common formats: GGUF files through llama.cpp on every platform, and MLX models on Apple Silicon Macs.
It is just as useful as plumbing for other software. Switch on the server and it answers on port 1234 in the docs' examples, speaking the OpenAI API, Anthropic's Messages API and its own REST API. Claude Code, Codex and most chat apps can use it by changing one base URL; LM Studio's own guide points Claude Code at http://localhost:1234.
Know what that server exposes: it is an open port on your machine. By default it binds to localhost (127.0.0.1), so only programs on the same computer can reach it, and it asks for no password, since authentication is off by default and needs version 0.4.0 or newer to switch on. Turning on Serve on Local Network, or starting it with lms server start --bind 0.0.0.0, opens it to every device on your network, and LM Studio itself recommends switching on API tokens before you do. MCP servers you add can run code and read your files; the docs warn against installing any from a source you do not trust.
For a shared box, version 0.4.0 brought llmster, the same engine with no window, installed with one command and started with lms daemon up. It also added parallel requests, 4 slots by default, so one long job no longer holds up everyone else's. LM Link, built with Tailscale, lets a laptop use models loaded on a bigger machine elsewhere over an end-to-end encrypted connection.
The catch: your hardware is the ceiling. On a Mac, LM Studio needs Apple Silicon and macOS 14.0 or later; Intel Macs are not supported, 16GB of memory is the recommended floor, and on Windows it suggests at least 4GB of dedicated graphics memory. The models that fit on such a machine trail Claude and GPT on hard tasks. The app is closed source, its terms allow personal and internal business use but forbid running it as a hosted service, and the agent features now sit in Bionic, a separate app with paid cloud plans at $20 and $100 a month.
Among the alternatives we review, Ollama does the same local serving as an open-source command-line tool, llama.cpp is the engine underneath both if you want full control, Jan is the open-source desktop equivalent, Magnitude tunes its GPU kernels to each model, and Unsloth's desktop app adds fine-tuning.
Our take: LM Studio is the local model app to hand a colleague who has never opened a terminal, and the one a developer can leave running as a private endpoint. Ollama wins for people who want open code end to end; LM Studio wins on browsing, comparing and loading models with a mouse.
Where it earns its place
A freelance developer on a 32GB Apple Silicon MacBook, working under client NDAs
Use a coding agent on code that cannot go to a cloud model
Load a model such as gpt-oss-20b, set ANTHROPIC_BASE_URL to http://localhost:1234 and run Claude Code against it, as LM Studio's guide shows. It costs nothing beyond the laptop, but a local model copes with far less than a frontier cloud model, so keep tasks small.
A six-person accounting practice
Summarise and query client documents without uploading them anywhere
Install the app on each 16GB laptop and chat with attached PDFs offline. Work use has needed no licence since July 2025, and the privacy policy says local chats and documents are never sent out. Answers from small models still need checking against the source.
An ML engineer looking after one shared Linux GPU server for a team of ten
Give the team an internal model endpoint
Install llmster, start it with lms daemon up, switch on token authentication and only then bind to 0.0.0.0, handing each colleague a named token. Parallel requests, 4 slots by default, keep one long job from blocking the rest.
Before you commit
- Setup
- Download the installer, pick a model from the built-in Hugging Face search and start chatting once it downloads; servers get llmster with one curl or PowerShell command.
- Integrations
- OpenAI-compatible /v1/chat/completions, /v1/responses and /v1/embeddings; Anthropic-compatible /v1/messages; its own REST API; the lmstudio-js and lmstudio-python SDKs; the lms CLI; MCP servers through a Cursor-style mcp.json; LM Link between your own devices.
- Data and privacy
- Local chats and documents are never transmitted from your system, according to the privacy policy, which also says the app has no telemetry. Update checks send the app version, operating system and IP address, and model searches are sent anonymised. Which AI tools train on your data?
- Cost at scale
- The software is free at any headcount; the real cost is hardware, since each person needs a machine with about 16GB of memory or access to a shared GPU server. Enterprise plans with SSO and model and MCP gating are sold through sales, with prices not published.
- Limits
- Server authentication is off by default; speed and context length depend on your memory and GPU; Intel Macs are unsupported, Windows on x64 needs AVX2, and the Linux AppImage is tested on Ubuntu 20.04 to 22.
What’s good
- Free for personal and work use, with no licence form since July 2025
- GGUF through llama.cpp everywhere, plus MLX on Apple Silicon
- OpenAI- and Anthropic-compatible server, REST API, TypeScript and Python SDKs
- llmster runs the same engine headless on Linux, Mac and Windows servers
- No telemetry, and local chats stay on the device, per the privacy policy
What’s not
- The server asks for no API token by default; network mode exposes it to the whole LAN
- The desktop app is closed source; only the CLI and SDKs are MIT-licensed
- Intel Macs are unsupported, and 16GB of memory is recommended
- Terms forbid offering it as a hosted service
- Agent features live in the separate Bionic app, with paid cloud plans
Best for
People who want private, offline AI on a Mac or PC without a terminal, and developers who need a local OpenAI- or Anthropic-compatible endpoint.
LM Studio alternatives
LM Studio: common questions
- Is LM Studio worth it?
- LM Studio is the friendliest way to run open models on your own computer, and its local server earns its keep well beyond the chat window, as long as you keep it on localhost. We rate it 4.4 out of 5.
- How much does LM Studio cost?
- LM Studio is free for personal and work use, with no licence form. The separate Bionic agent app has a free plan, Bionic+ at $20/month and Pro at $100/month for hosted open models; Teams and Enterprise prices are not published.
- What are the downsides of LM Studio?
- The server asks for no API token by default; network mode exposes it to the whole LAN; the desktop app is closed source; only the CLI and SDKs are MIT-licensed; intel Macs are unsupported, and 16GB of memory is recommended; terms forbid offering it as a hosted service; agent features live in the separate Bionic app, with paid cloud plans.
- Who is LM Studio best for?
- People who want private, offline AI on a Mac or PC without a terminal, and developers who need a local OpenAI- or Anthropic-compatible endpoint.
- How much does LM Studio cost for a team or in production?
- The software is free at any headcount; the real cost is hardware, since each person needs a machine with about 16GB of memory or access to a shared GPU server. Enterprise plans with SSO and model and MCP gating are sold through sales, with prices not published.
- Is LM Studio safe for company data?
- Local chats and documents are never transmitted from your system, according to the privacy policy, which also says the app has no telemetry. Update checks send the app version, operating system and IP address, and model searches are sent anonymised.
- What are the best LM Studio alternatives?
- The closest alternatives we have reviewed are Local Code.
Answers checked Nov 14, 2026.
Make LM Studio? Show this rating on your site
