LM Studio: Local LLMs

What is LM Studio?

LM Studio is a local LLMs platform that lets you download and run AI language models directly on your own computer. It covers a wide range of open-weight models and gives you full control over how they run. Internet access is not required once models are downloaded. You can interact through a desktop app with a built-in chat interface, use a headless daemon for servers and CI environments, or access models through a REST API. Local LLMs work on Mac, Windows, and Linux through LM Studio, with Apple Silicon Macs also supported via Apple’s MLX framework. A JavaScript SDK and Python SDK are available for building custom apps and scripts on top of local LLMs. A CLI tool handles model downloads, server control, and daemon management from the terminal. An iPhone app lets you run models on the go.

Features & Benefits

• Local model download and management: search, download, and manage open-weight LLM files directly within the app, sourced from Hugging Face.
• Chat interface: run interactive conversations with local LLMs through a built-in desktop chat UI with configurable settings per model.
• Document chat (RAG): attach documents to chat sessions and query them with local LLMs entirely offline, with no data leaving your machine.
• Headless deployment (llmster): deploy and run local LLMs on servers, cloud instances, or CI pipelines without a GUI using the llmster daemon.
• Local REST API: serve local LLMs over HTTP with a REST API that supports stateful chats, model loading, downloading, and streaming responses.
• OpenAI-compatible endpoints: access local LLMs through OpenAI-style endpoints covering chat completions, embeddings, structured output, tool use, and responses.
• Anthropic-compatible endpoints: call local LLMs using Claude-style Messages API flows against a local LM Studio server.
• JavaScript SDK: build local LLM workflows, tools, and apps using the lmstudio-js TypeScript/Node.js SDK.
• Python SDK: run local LLMs from Python scripts, notebooks, and backend services using the lmstudio-python package.
• CLI (lms): manage local LLMs from the terminal with commands for model downloads, daemon control, server startup, and interactive chat.
• MCP client support: connect MCP servers to LM Studio and use them with local models for tool-calling and agent workflows.
• Speculative decoding: increase local LLM generation speed using speculative decoding with a draft model.
• Parallel requests: handle multiple simultaneous inference requests from a single local model instance.
• Per-model configuration: set default parameters and prompt templates on a per-model basis for consistent local LLM behavior.
• Config presets: save, import, export, and publish reusable model configuration presets.
• LM Link: route local LLM workloads across multiple devices and preferred machines on a network.
• GGUF and MLX model support: run GGUF models via llama.cpp on all platforms and MLX models on Apple Silicon Macs.
• Structured output: generate JSON outputs that validate against a defined schema when calling local LLMs via API.
• Embeddings: produce vector embeddings from local LLMs for retrieval, search, and indexing pipelines.
• Offline operation: run all local LLM features fully offline after initial model download.

What can LM Studio do?

• Run LLMs locally on your computer
• Chat with documents offline using local AI
• Deploy local LLMs on a server without a GUI
• Serve local models on an OpenAI-compatible API
• Use local LLMs with MCP tool-calling
• Generate structured JSON output with a local model
• Create text embeddings with a local LLM
• Download open-weight models from Hugging Face
• Build apps with a local LLM using Python or JavaScript
• Run AI models privately without sending data to the cloud
• Use local AI models on Apple Silicon with MLX
• Run local LLMs in a CI pipeline

Real-World Applications

Privacy-sensitive work is a natural fit for local LLMs. Legal professionals, healthcare workers, and finance teams dealing with confidential documents may find it useful to run AI on their own hardware. Sensitive text never leaves the machine. Document chat through local RAG means contracts, reports, or case notes can be queried directly without connecting to any external service.

Developers building AI-powered apps can use local LLMs as a drop-in backend during development. The OpenAI-compatible and Anthropic-compatible endpoints mean existing code often works with minimal changes. Python and JavaScript SDKs make it straightforward to prototype local LLM features, test prompt logic, and iterate quickly without per-call API costs.

Teams running automated pipelines can deploy local LLMs in CI environments or on dedicated servers through the headless daemon. A build process that needs text generation, classification, or structured output can call local LLMs the same way it calls any local service. This keeps AI inference internal and reproducible across environments.

Researchers and technically minded individuals who want to experiment with open-weight models can use local LLMs without cloud accounts or usage caps. Trying different models, testing prompt templates, and benchmarking outputs is possible entirely on local hardware. The ability to run multiple models and switch between them makes systematic comparison practical.