LM Studio is a free desktop application for Windows, macOS, and Linux that downloads and runs open-weight large language models on your computer. It combines a model catalog, a chatbot, document chat, model controls, and a programmable local server in one graphical application.
The app runs GGUF models through llama.cpp on supported desktop systems. Apple Silicon Macs can also run MLX models. After you download the model and its runtime, chats, document processing, and local API requests can stay on your device and work offline.
Model search, downloads, runtime downloads, and app updates need an internet connection.
LM Studio also recently introduced LM Studio Bionic, the next-generation agent for research, document, and code tasks. It can work on files with local, remote, or cloud models. If you want to download and run a model directly, install the classic app.
LM Studio vs LM Studio Bionic
| LM Studio | LM Studio Bionic |
|---|---|
| Desktop app for downloading, loading, chatting with, and serving local models | Next-generation agent for document, research, and coding projects |
| Runs downloaded models locally through llama.cpp or MLX | Uses local models, LM Link remote models, or LM Studio cloud models |
| Provides native REST, OpenAI-compatible, and Anthropic-compatible endpoints | Works through project sessions that can create or modify files |
Key Features
- Search Hugging Face by model family, repository name, or full URL.
- Keep conversations in folders, duplicate a thread worth saving, or open two chats side by side.
- Attach PDF, DOCX, and TXT files as local context through retrieval-augmented generation.
- Adjust context length, sampling, prompt templates, GPU offload, and per-model defaults.
- Serve a loaded model through LM Studio’s native REST API, OpenAI-compatible endpoints, or Anthropic-compatible endpoints.
- Manage models and the server with the
lmsCLI. TypeScript and Python SDKs connect local applications to the same runtime. - Connect local or remote MCP servers and expose their capabilities to a compatible local model.
- Run downloaded models, document chat, and local API requests offline after the required model and runtime files are present.
Use Cases
- Private chat sessions keep prompts on a personal computer.
- Local document retrieval can examine private reports, manuals, notes, and research papers.
- OpenAI-compatible and Anthropic-compatible requests can power a local application prototype.
- Coding clients connect to the local server through a custom base URL.
- Test tool calling, structured JSON output, embeddings, tokenization, and MCP connections on your own machine.
- One well-equipped desktop can serve inference to trusted devices on the local network.
- An offline AI workstation fits travel, restricted networks, and confidential material.
How to use it
1. Install the classic LM Studio app
Open the LM Studio download page and select the classic application for Windows, macOS, or Linux. Run the installer and open the app. Local models do not need a hosted-model API key or an LM Studio cloud account.
2. Install a runtime and download a model
LM Studio uses separately packaged runtimes to load different model formats. Press Command + Shift + R on macOS or Ctrl + Shift + R on Windows and Linux to inspect installed runtimes. Open Discover, choose a model that fits the available memory, and download a 4-bit or higher variant for the first test.
3. Load the model and start a first chat
Open Chat, select the downloaded model from the model loader, and keep the default load settings for the first run. Send a short prompt that has an obvious answer. This confirms that the runtime, model file, and memory settings work before you raise the context length or change sampling controls.
4. Chat with local documents
Attach a PDF, DOCX, or TXT file to a chat message. LM Studio retrieves relevant passages and supplies them to the loaded model as context. The document processing stays on the computer during local use. Check the answer against the source file because retrieval can miss a section and the model can misread it.
5. Tune the model after the baseline works
Change context length, temperature, GPU offload, or the prompt template one setting at a time. Save a stable configuration as a preset before another round of experiments. Developer Mode under Settings > Developer exposes the complete set of advanced controls and development features.
6. Start the local model server
Start the server inside LM Studio or run lms server start. The default address is http://localhost:1234. Your application connects through LM Studio’s native REST API, an OpenAI-compatible client, or an Anthropic-compatible client. Bind the server to the local network only when other trusted devices need access, and use authentication before exposing it beyond your own machine.
7. Use the CLI for repeatable tasks
The lms command-line interface is installed with the desktop app. It can download and load models, control the server, open an interactive terminal chat, and stream request logs.
| Command | Purpose |
|---|---|
lms get <model> | Search for and download a model. |
lms load <model> | Load a downloaded model into memory. |
lms ls | List models stored on the machine. |
lms server start | Start the local HTTP server. |
lms chat | Open an interactive terminal chat. |
lms log stream | Watch incoming and outgoing request logs. |
8. Review MCP permissions before connecting a server
Open the Program tab, choose Install > Edit mcp.json, and add a local or remote MCP server using the supported configuration format.
9. Prepare the machine for offline use
Download the application, model files, and required runtime before the computer goes offline. Test the exact chat, document, or API task once with the network disconnected.
How to download compatible model files from Hugging Face
LM Studio searches supported Hugging Face models inside the Discover tab. Search by model family, a specific owner/model name, or a full Hugging Face URL. The old habit of adding ggml to every query is obsolete.
- Open Discover. Press
Command + 2on macOS orCtrl + 2on Windows and Linux to jump there. - Enter a model name such as
Qwen,Llama, orgpt-oss. You can paste the complete Hugging Face model URL when you already know the repository. - Pick the model format that matches your runtime. GGUF is the normal llama.cpp choice. MLX is available on Apple Silicon Macs.
- Select a quantization. Names such as
Q3_K_S,Q4_K_M, andQ8_0describe different size and fidelity tradeoffs. A 4-bit option or higher is a sensible first download when the hardware can hold it. - Download the file, load it in Chat, and test a modest context length before raising memory-heavy settings.
Which configuration parameters should I choose?
Use the model’s default preset for the first successful chat. Change one control at a time after you have a baseline. Random changes across temperature, sampling, prompt format, context length, and hardware settings make a bad result harder to diagnose.
Adjusting configuration parameters
- Temperature changes output randomness. Lower values tend to produce more repeatable answers. Higher values produce more variation and more risk of drift.
- Top P restricts token selection to a probability mass. Leave it at the preset value until you have a reason to tune sampling.
- Context length controls how much text the model can consider in one request. Longer contexts consume more RAM or VRAM, especially through the key-value cache.
- System prompt sets the model’s role or persistent instructions for the active conversation. Clear, narrow instructions beat a page of vague rules.
- Prompt template must match the model’s expected chat format. LM Studio usually reads this from the model metadata. Override it only when the model card calls for a specific template.
- Hardware and load settings control GPU offload, CPU threads, and memory use. Start with automatic settings. Manual tuning earns its keep when a model loads slowly, spills out of VRAM, or fails with an out-of-memory error.
- Presets save a working collection of inference and prompt settings. Export a preset before major experiments if you want an easy route back to a known result.
Which model size should I choose? (7B, 13B, 15B, 30B, etc.)
The old rule that maps one RAM number to one parameter count is too crude to trust. Quantization, context length, runtime, GPU offload, and architecture all change the memory bill. Two 7B models can behave very differently after one is quantized to 4-bit and the other is loaded at higher precision with a long context.
| Check | Practical decision |
|---|---|
| Model download size | Leave extra memory for the runtime, context cache, and operating system. |
| Quantization | Start at 4-bit or higher, then compare quality and speed. |
| Context length | Begin modestly; long contexts can consume several additional gigabytes. |
| RAM and VRAM | Use the catalog’s device-fit signal and avoid filling either pool completely. |
| First test | Load a small model, confirm stable generation, then move upward. |
LM Studio recommends 16GB or more of system memory on Windows and macOS, although smaller models can run on an 8GB Mac with modest context settings. A Windows x64 processor needs AVX2, and 4GB of dedicated VRAM is the recommended floor for GPU use. Bigger parameter counts do not guarantee a better daily model. A smaller model that fits cleanly and answers quickly is often the one you will actually keep running.
Pros
- Local chats can run fully offline.
- Model discovery connects directly to Hugging Face.
- Local APIs fit existing development clients.
- GGUF runs across supported desktop platforms.
- MLX runs on Apple Silicon Macs.
Cons
- Large models demand substantial memory.
- Model downloads require internet access.
- Untrusted MCP servers can execute code.
FAQs
Q: Does LM Studio run on Windows 11?
Yes. Use the x64 build on a conventional Intel or AMD Windows 11 PC. Snapdragon-based Windows 11 devices need the ARM64 build.
Q: Can I download LM Studio for an Intel Mac?
No. LM Studio releases require an Apple Silicon Mac. The supported macOS baseline is macOS 14, and 16GB of RAM is recommended.
Q: Is the LM Studio desktop app open source?
No. Element Labs distributes LM Studio under its own software terms, and the desktop application’s source code is proprietary. The models and runtimes loaded through LM Studio have separate licenses. Check each model card before commercial use or redistribution.
Q: Can LM Studio run a model with CPU only?
Yes. The llama.cpp runtime can run models on the CPU, and GPU use is optional. CPU-only generation is usually slower. Pick a small quantized model and a modest context length when the machine has limited memory.
Q: Can I import a GGUF file downloaded outside LM Studio?
Yes. Run lms import <path/to/model.gguf> and follow the interactive prompt. LM Studio expects imported files under a publisher and model directory inside ~/.lmstudio/models/.
Q: Can I move the LM Studio models folder to another drive?
Yes. Open My Models and change the models directory. Moving a large library to a faster or larger drive can free the system disk, but LM Studio must index the selected directory before those files appear.
Q: Where does LM Studio store chat history?
Right-click a chat and select Reveal in Finder or Show in File Explorer. Conversations are JSON files under ~/.lmstudio/conversations/ on macOS and Linux, or %USERPROFILE%\.lmstudio\conversations\ on Windows.
Q: Does a local model learn from LM Studio chats?
No. A model uses the messages and system prompt included in its active context, but a normal chat does not retrain the model or change its weights. A new conversation has no memory of an earlier thread unless you supply that information again.
Q: What is llmster, and do I need the LM Studio interface to use it?
llmster is LM Studio’s standalone headless daemon. It runs model-serving workloads on a server, a GPU machine with no display, a CI environment, or a background service. It does not depend on the LM Studio desktop interface, and the lms CLI controls it.
Q: Does LM Studio Bionic require a paid plan?
Bionic’s local model use is free, and local voice transcription runs on your own machine. LM Link models also avoid cloud credits. Hosted cloud models require an LM Studio account, billing setup, and available credits.
Alternatives and Related Resources
- Nativ: Run Local AI Models Natively on Macs
- SmallCode: Fast, Free, Local AI Coding Agent for Small LLMs
