cloudparts

LM Studio

A free, closed-source desktop app and headless daemon that downloads open-weight models from Hugging Face, runs them on your own machine with llama.cpp or Apple’s MLX, and serves them on OpenAI- and Anthropic-compatible endpoints at localhost:1234.

What it is
Running models locally
Type
Commercial product
Made by
Element Labs, Inc.
License
Proprietary, free for personal and work use; the lms CLI, the lmstudio-js and lmstudio-python SDKs and the MLX engine are MIT
Facts checked
13 October 2026
Website

What it is

LM Studio searches Hugging Face for builds of open models, downloads the weights to disk and runs them with an LM Runtime: llama.cpp (GGUF files) on macOS, Windows and Linux, and Apple’s MLX on Apple Silicon. Runtimes update separately from the app. Once a model is on disk the app works offline: chats, attached documents and the local server stay on the machine, and only model search, downloads, runtime downloads and update checks reach the network. A document short enough to fit the model’s context is added whole; a longer one goes through built-in retrieval (RAG). It runs on Apple Silicon Macs with macOS 14 or newer, x64 and ARM Windows, and x64 and ARM64 Linux as an AppImage.

The Developer tab or lms server start opens an HTTP server on port 1234 with LM Studio’s own REST API, OpenAI-compatible endpoints (/v1/chat/completions, /v1/responses, /v1/embeddings) and, since 0.4.1, an Anthropic-compatible /v1/messages, so existing SDKs, Codex and Claude Code reach local models by changing a base URL. A model loads just in time on its first request and unloads after an idle TTL, 60 minutes by default; with Auto-Evict on, at most one JIT-loaded model stays in memory. Version 0.4.0 (January 2026) split the core from the GUI as llmster, a daemon for Linux servers, GPU rigs without a display and CI, and added parallel requests with continuous batching to the llama.cpp engine.

The app itself is closed source but free, and since July 2025 that includes use at work without a separate commercial license. Element Labs sells an Enterprise plan with SSO and model and MCP gating. In July 2026 it also released LM Studio Bionic, a separate agent app that runs local models through the LM Studio runtime or larger open models on its billed LM Studio Secure Cloud.

Why teams pick it

  • Prompts, documents and weights stay on your machine once a model is downloaded, and the app works with no network connection
  • Two engines: llama.cpp on every platform and MLX on Apple Silicon, where Element Labs measured Llama 3.2 1B at about 250 tokens per second on an M3 Max when MLX support shipped
  • Drop-in local endpoints for OpenAI and Anthropic clients, with just-in-time loading, idle TTL and auto-evict so editors and agents can switch models without manual loading
  • Runs headless as llmster on Linux servers and in CI, driven by the same lms CLI as the desktop app

Watch out for

  • The app is closed source; only the CLI, SDKs and MLX engine are open (MIT), so the core can’t be audited or forked
  • Model size and speed are bounded by your hardware: the docs recommend 16 GB of RAM and, on Windows, at least 4 GB of dedicated VRAM, and say 8 GB Macs should stick to small models and modest context sizes
  • No Intel Macs, x64 Windows PCs need AVX2, and the Linux build is not well tested on Ubuntu releases newer than 22
  • Continuous batching arrived in 0.4.0 for the llama.cpp engine only; Element Labs said the MLX engine would follow

Alternatives

What all running models locally have in common →Compare side by side →
AlpacaOpen sourceA GNOME desktop app for Linux that chats with local models through Ollama, which it can install and run for you, and with cloud models through OpenAI-compatible APIs using your own keys.Open WebUICommercialA self-hosted web app for chatting with models served by Ollama or any OpenAI-compatible API, with users and groups, document retrieval, tools and web search, under a source-available license that requires keeping its branding above 50 users.

How it fits with other technologies

Install the headless daemon, download a model and serve it on localhost:1234 · bash

curl -fsSL https://lmstudio.ai/install.sh | bash   # llmster; the desktop app ships lms too

lms daemon up
lms get openai/gpt-oss-20b
lms load openai/gpt-oss-20b --ttl 3600             # unload after an idle hour
lms server start

curl http://localhost:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "openai/gpt-oss-20b",
       "messages": [{"role": "user", "content": "Say this is a test!"}]}'

See it in a real stack

No architecture uses LM Studio yet. Drop it into the playground and wire it up yourself.Open playground →
Where it came from2023–2026

LM Studio launched in May 2023 as a desktop app for running open models on a personal computer. Its maker, Element Labs, describes the premise as AI that is accessible to people on their own machines, “with no reliance on an external party, and with complete privacy.” It stayed a small team: nine people in July 2025.

The app grew from a chat window into a local AI server. In May 2024 it gained lms, an MIT-licensed command-line tool developed on GitHub. In October 2024 version 0.3.4 added an MLX engine for Apple Silicon beside llama.cpp. In July 2025 Element Labs dropped the rule that work use needed a separate commercial license, after seeing teams avoid the app rather than start a procurement process. Version 0.4.0 in January 2026 separated the core from the GUI as the llmster daemon and added parallel requests, and 0.4.1 added an Anthropic-compatible endpoint so Claude Code could use local models.

In July 2026 Element Labs launched LM Studio Bionic, a separate agent app for coding and document work. It runs local models through the LM Studio runtime and larger open models on LM Studio Secure Cloud, which Element Labs says keeps no data after a request.

Why it’s built this way

Run open models on the user’s own machine
The team believes AI should be usable on your own computer without relying on an outside party, and with complete privacy.
Swappable runtimes: llama.cpp plus MLX
MLX runs on-device models efficiently on Apple Silicon, and packaging engines as runtimes lets them update without a full app release.
Free at work, paid Enterprise features
Requiring a commercial license for work use made adoption at companies high-friction, so Element Labs charges only for team needs such as SSO and model or MCP gating.
Split the core from the GUI
To run the same model serving on Linux servers, GPU rigs without a screen, CI pipelines and Colab.

Timeline

  1. 2023

    First release of the desktop app (May)

  2. 2024

    lms CLI released under the MIT license (May)

  3. 2024

    MLX engine for Apple Silicon in 0.3.4 (October)

  4. 2025

    Free for use at work (8 July)

  5. 2026

    0.4.0 adds the llmster daemon and parallel requests; 0.4.1 adds an Anthropic-compatible API (January)

  6. 2026

    LM Studio Bionic agent app with LM Studio Secure Cloud (16 July)

Sources13 · facts checked 13 October 2026

Improve this page on GitHub