LM Studio is a desktop application for downloading open language models and running them on your own computer. It combines a chat window, a model browser and a local server in one package, and on its current home page the company also promotes a newer feature called Bionic, an agent built for open models. This is a “what it is and who it suits” overview based on the company’s own site and documentation. We have not run a hands-on test, so we make no claims about speed or quality on any particular machine.

What LM Studio is

According to its home page, LM Studio is built around running models “natively local”: you download open models inside the app and use them for simple chats or more involved agentic work. The site says the app is powered by the LM Studio runtime, with MLX and llama.cpp under the hood. MLX is Apple’s machine-learning framework for Apple Silicon, and llama.cpp is a widely used open-source inference engine, so the app is essentially a friendly shell around engines that many local-AI users already know.

The company behind the product is listed in the site footer as Element Labs, Inc. The documentation also describes related pieces: a command-line tool called lms, a headless mode (llmster) for running the engine as a service, and client libraries for JavaScript and Python (lmstudio-js and lmstudio-python).

Bionic, the agent layer

The home page currently leads with Bionic, described as LM Studio’s agent for open models. The site says Bionic can create and edit documents, with every change saved automatically, and handle coding tasks, automations and computer control. It also lists real-time voice transcription that the company says is processed locally and never leaves the device.

The site names frontier open models such as GLM 5.2, Kimi K3 and DeepSeek V4 Pro as options for demanding tasks. Running models of that class locally generally needs serious hardware, so treat those as the upper end rather than the typical laptop experience. The company also says its cloud services are Zero Data Retention. That is a vendor statement we cannot verify independently.

Running a local API server

For developers, the most practical feature is the local server. The documentation explains that you can serve models from the Developer tab, either on localhost or on your local network, by toggling “Start server” or by running lms server start in a terminal. The documented API options are:

  • a native REST API with endpoints to list, load, unload and download models and to chat, including streaming and stateful chats;
  • OpenAI-compatible endpoints, including chat completions, embeddings, responses, structured output and tool use;
  • Anthropic-compatible endpoints for the Messages API;
  • compatibility with the Jev (TypeSafe AI) decision API;
  • MCP support via the API, plus idle time-to-live and auto-evict settings that unload models you are not using.

In practice, compatibility endpoints mean that code written against a hosted provider can often be pointed at a local address with a changed base URL. Whether a particular feature behaves identically depends on the model you load, so test your own workload. Our guide to function calling and structured outputs explains why those features vary by model.

System requirements

The documentation lists the following. On macOS you need an Apple Silicon chip (M1 through M4) and macOS 14.0 or newer; 16 GB or more of RAM is recommended, and 8 GB Macs may work with smaller models and modest context sizes. Intel Macs are not supported. On Windows, both x64 and ARM (Snapdragon X Elite) systems are supported; x64 needs the AVX2 instruction set, at least 16 GB of RAM is recommended, and at least 4 GB of dedicated GPU memory is suggested. On Linux, x64 and ARM64 are supported, distributed as an AppImage, with Ubuntu 20.04 or newer required and newer-than-22 versions described as less thoroughly tested. The docs also say LM Studio can operate entirely offline once you have downloaded model files. For sizing models to your memory, see our GPU memory guide.

LM Studio compared with Ollama

LM Studio and Ollama overlap heavily: both run open models locally and offer a local API. The difference is mostly in emphasis. LM Studio leads with a graphical app, a built-in model browser and a growing agent product, while Ollama is better known for a lightweight command-line workflow; see our earlier Tool of the Day on Ollama and the step-by-step Ollama guide. Many people install both. Pick on workflow rather than on assumed performance, since both rely on similar underlying engines.

Who it suits

LM Studio looks like a good fit for people who want a point-and-click way to try open models, for developers who need a local endpoint for prototyping without per-token charges, and for anyone handling material they prefer not to send to a cloud service. It is less obviously suited to teams that need managed multi-user serving at scale, where dedicated inference servers are the usual choice. The trade-offs of going local are covered in our comparison of cloud APIs and local models.

What to watch next

  • How the Bionic agent evolves and which open models it supports well.
  • Changes to the documented API compatibility, including the Anthropic and Jev endpoints.
  • Whether Intel Mac or other currently unsupported platforms are added.
  • Independent benchmarks comparing local runtimes on the same hardware.
  • Any changes to licensing or pricing terms, which you should confirm on the company’s own pricing page.

Key takeaways

  • LM Studio is a desktop app for downloading and running open models locally, with chat, a model browser and a server.
  • Its documentation lists OpenAI-compatible and Anthropic-compatible endpoints plus a native REST API and CLI.
  • It runs on Apple Silicon Macs, Windows (x64 and ARM) and Linux, with 16 GB of RAM recommended.
  • The site promotes an agent called Bionic; some claims, such as zero data retention for cloud services, are vendor statements.
  • This overview is based on public documentation, not hands-on testing.

Frequently asked questions

Does LM Studio work without an internet connection?

The documentation says it can operate entirely offline, as long as you have already downloaded the model files you want to use.

Can I use it with code written for the OpenAI API?

The docs describe OpenAI-compatible endpoints, so many clients can be pointed at the local server. Check the specific features you rely on, such as tool use, against the model you load.

Is it free?

We have not verified current pricing, so check the company’s pricing page before relying on any plan or license terms.