LM Studio

General definition

LM Studio is a desktop application for running open-weight language models on your own computer without touching a terminal. It has a built-in model browser that searches Hugging Face, downloads a model in GGUF format (or MLX format on Apple Silicon), shows whether it will fit in your memory, and opens a chat window the moment the download finishes. It uses the llama.cpp engine for GGUF models and Apple’s MLX framework on Macs.

For developers the important feature is the local server. One click (or lms server start from the lms command-line tool) exposes the loaded model on http://localhost:1234/v1 with OpenAI-style endpoints for chat completions, completions, embeddings and model listing. Any application that already speaks the OpenAI-compatible API can switch to the local model by changing its base URL. The server supports streaming, tool calling and structured output, can keep several models loaded at once, and runs headless on a workstation with the window closed.

• Free to download and use, including at work since 2025; the application itself is proprietary, while the lms CLI and the Python and TypeScript SDKs are open source
• Model controls in the interface: context length, GPU offload, sampling settings, system prompt and quantisation choice
• Chat with local documents inside the app, and a model catalogue curated by the LM Studio team

Where it fits: LM Studio is built for one person on one machine. It is the fastest way for a developer to try a model, compare quantisations, or run a pilot of an AI feature against a local LLM endpoint. Ollama covers the same ground with a command-line, scriptable, server-friendly design, and vLLM is the step up when many users need to share one GPU server. A small language model in LM Studio on a laptop is often all a prototype needs.

In the Ethora ecosystem

Ethora’s AI service connects to any OpenAI-compatible endpoint, and LM Studio’s local server is one. The same four environment variables used in the official Ollama example (AI_API_URL, AI_API_KEY, AI_CHAT_MODEL, AI_EMBEDDING_MODEL) point the self-hosted AI agent stack at http://localhost:1234/v1 and the models you have loaded. For RAG the knowledge base uses Postgres with pgvector and 256-dimension embeddings, so load an embedding model in LM Studio as well as a chat model. Ethora publishes a step-by-step tutorial for this setup on its blog.

This is the single-developer path: run an AI agent in a chat room on your laptop, compare a few models, demo to stakeholders, all without an API key or any outbound traffic. The endpoint is configured once for the whole service (agents choose a model name, not an endpoint), so when the pilot grows you swap the LM Studio URL for Ollama on a server or a vLLM cluster, and the agents, prompts and knowledge base stay exactly as they are.