Source: LM Studio documentation (app and developer docs)

LM Studio's docs: a desktop app running GGUF via llama.cpp and MLX models on Apple silicon, with OpenAI- and Anthropic-compatible local server, lms CLI, SDKs and JIT loading with TTL.

Contents

Summary

LM Studio is a desktop GUI for finding, downloading and chatting with local models. It also has a developer surface: a local HTTP server, the lms CLI, a headless daemon, and JavaScript and Python SDKs.

Key claims

  • Runs GGUF models with llama.cpp on macOS, Windows and Linux. On Apple silicon it also runs MLX models.
  • System requirements: macOS 14.0 or newer on Apple silicon only (Intel Macs unsupported); Windows on x64 (AVX2 required) or ARM (e.g. Snapdragon X Elite); Linux on x64 or ARM64 as an AppImage (Ubuntu 20.04 or newer).
  • Works fully offline once the models are downloaded, including chat with attached documents (local RAG).
  • The local server (lms server start, conventionally on port 1234) exposes OpenAI-compatible endpoints (Chat Completions, Responses, Embeddings, Models, legacy Completions) with streaming, tool use and JSON-schema structured output. It also offers an Anthropic-compatible Messages endpoint and its own REST API.
  • SDKs are lmstudio-js (@lmstudio/sdk) and lmstudio-python (lmstudio). llmster packages the core as a headless daemon for servers and CI.
  • Just-in-time loading brings a model into memory on its first request. JIT-loaded models unload after an idle TTL, 60 minutes by default, and a per-request ttl field in seconds overrides it. Auto-evict, on by default, keeps only one JIT-loaded model resident at a time.

Feeds

Local LLMs

Suggest changes →