LM Studio

LM Studio is a desktop application that handles every step of running a model locally through a graphical interface: search for a model inside the app, see which version fits your computer, click to download, load it and chat in the built-in window. You never have to open a terminal, which makes it friendly for people who are not comfortable with the command line.

Compared with Ollama, its advantage is visibility. When you search, it shows the size of each quantised version and whether your hardware can handle it. When you load a model, you can adjust context length, GPU offload and other parameters directly. The chat interface is fuller, and you can drop documents in and ask about them. On Apple silicon Macs it also supports Apple’s MLX framework, which runs efficiently.

It can also start a local server with an OpenAI-compatible API for other tools to use. Note that LM Studio is free but not open source, which matters to some users. And its built-in model search and downloads rely on Hugging Face, which usually cannot be reached directly from mainland China, so users there need to supply model files themselves or find another route.

Key features

• Find models in the app: search open models and see each quantised version’s size with hints about hardware fit.
• Chat interface: load a model and chat in a graphical UI, with multiple conversations and system prompts.
• Chat with documents: drop local documents into a conversation and have the model answer from them.
• Visual parameter controls: context length, GPU offload, sampling settings and more, all adjustable in the interface.
• Local API server: switch on an OpenAI-compatible local endpoint for other applications.
• Multiple inference backends: supports common quantised model formats, with MLX acceleration on Apple silicon.

How to use

1. Download the installer for your system from the website and install it.
2. Search for a model in the app, preferring a quantised version marked as suitable for your hardware.
3. Once downloaded, load it on the chat page, start talking and adjust parameters as needed.
4. To serve a model to other tools, start the local server from the developer page and copy the endpoint address.
5. If in-app downloads fail, download model files from a domestic model hub and import them into LM Studio’s models folder.

Best for

• Local-model users who want to avoid the command line: graphical from start to finish.
• Choosing a model that fits your computer: hardware hints save a lot of trial and error.
• Private document Q&A: the model reads your files on your machine, and nothing leaves it.
• Serving a model backend to other tools: a local endpoint for plugins and clients.

Strengths and limitations

Strengths

• A complete graphical interface; no command line needed from finding a model to chatting.
• Clear hardware hints and parameter controls, good for beginners finding their way.
• Runs efficiently on Apple silicon Macs.

Limitations

• The app is not open source, which rules it out where open source is required.
• Built-in downloads rely on Hugging Face, which usually cannot be reached from mainland China.
• Local model quality is limited by hardware and falls well short of cloud flagships.
• Heavier on resources than Ollama when left running as a background service.