General · EdgepediaSeptember 17, 20267 min read

LM Studio

LM Studio is a desktop application for discovering, downloading, and running large language models locally on a personal computer, developed by Element Labs Inc., a software startup based in New York, USA.¹ The app provides a graphical chat interface, a built-in model downloader for Hugging Face, and a local API server that exposes OpenAI-compatible, Anthropic-compatible, and native REST endpoints, so the same models can be used interactively or from code.²^•³ It runs on Apple Silicon Macs, Windows, and Linux, and can operate entirely offline once models are downloaded.⁴^•⁵

Key fact	Detail
Developer	Element Labs Inc., New York, USA¹
First released	May 2023, initially free for personal use with a separate commercial license required for work⁶
Licensing change	Free for use at work since July 8, 2025, with no form or license request needed⁶
Inference engines	llama.cpp-based engine plus an MLX engine for Apple Silicon, shipped in version 0.3.4 (October 8, 2024)⁷
APIs	OpenAI-compatible, Anthropic-compatible, and native REST endpoints; TypeScript and Python SDKs (lmstudio-js, lmstudio-python)³^•⁸
Privacy	Local chats, documents, and histories are never transmitted; the developer receives data only for update checks, model search and download, and optional cloud features⁹
Requirements	Apple Silicon (M1–M4) with macOS 14.0+, 16 GB RAM recommended; Windows x64/ARM with AVX2; Linux x64/ARM64 AppImage⁴

Purpose and interface

The application is designed to let people run open models on their own hardware without sending data to an external service. Its main window is a ChatGPT-like chat interface in which users create conversation threads, organize them into folders, and view them side by side in a split view; conversations are stored locally as JSON files under ~/.lmstudio/conversations/ on Mac and Linux, or %USERPROFILE%\.lmstudio\conversations on Windows.¹⁰

Model discovery and download. A built-in downloader searches Hugging Face by keyword, by a user/model string, or by full URL, and offers the same model in several quantization levels, where the Q labels (such as Q3_K_S or Q8) indicate how aggressively the model files are compressed at some cost in quality; the documentation advises choosing a 4-bit option or higher when the machine can run it.² Models downloaded outside the app can be imported with lms import or by placing files in the expected ~/.lmstudio/models/publisher/model/ directory structure.¹¹

Documents. Users can attach .docx, .pdf, and .txt files to chat sessions. If a document fits the model's context window, its contents are added in full; if it is very long, LM Studio uses retrieval-augmented generation (RAG) to select relevant portions as context. All document processing happens locally.¹²^•⁵

Inference engines

LM Studio supports multiple inference engines. The original engine is based on llama.cpp, the project that also underlies comparable tools such as Ollama, and it handles GGUF model files.¹³ In October 2024, version 0.3.4 added an MLX engine for Apple Silicon Macs, reporting about 250 tokens per second for Llama 3.2 1B on an M3 Max, with JSON-constrained output provided through the Outlines library.⁷ In May 2025 the MLX engine was rearchitected into a unified multi-modal design built on mlx-lm and mlx-vlm, and its source is available in the company's open-source mlx-engine repository.¹⁴

Load-time settings can be tuned per model, including GPU offload, context length, and Flash Attention, and saved as defaults.¹⁵ The REST load endpoint exposes the same controls, including offload_kv_cache_to_gpu, eval_batch_size, and the number of experts used at inference for Mixture-of-Experts models.¹⁶ Version 0.3.10 (February 2025) added speculative decoding for both engines, and version 0.3.14 (March 2025) added multi-GPU controls for enabling or disabling specific GPUs and choosing allocation strategies.¹⁷

Local server and developer tooling

From the Developer tab, or with the command lms server start, LM Studio runs an API server on localhost that serves loaded models over HTTP.³ Clients can connect through the native REST API, the TypeScript SDK lmstudio-js, the Python SDK lmstudio-python, or compatibility endpoints that mimic the OpenAI and Anthropic APIs, which lets tools built for those services, such as Claude Code, run against local models.³^•¹⁷ The SDKs reached a 1.0.0 release on March 3, 2025 and expose the same capabilities as the desktop app, including precise control over loading and unloading models from memory.⁸^•¹⁸

The lms command-line tool, introduced in May 2024, handles model downloads, the background daemon, server startup, and scripted workflows; version 0.3.5 (October 2024) added headless mode, on-demand model loading, and server auto-start so the app can act as a background LLM provider without its window open.¹⁷^•¹⁹ The server can be bound beyond localhost with lms server start --bind 0.0.0.0 to serve other devices on the same network, a configuration the documentation flags as requiring authentication.²⁰ API token authentication, CORS control, and just-in-time model loading with automatic unloading are configured in Server Settings.²¹^•²²

Tool use and MCP. Version 0.3.6 (January 2025) introduced a beta tool-calling API, and version 0.3.17 (June 2025) added Model Context Protocol (MCP) host support, allowing local models to call external tools through configured MCP servers.¹⁷ From version 0.4.0, MCP servers can also be invoked through the API, either as ephemeral per-request servers or as servers pre-configured in mcp.json, with an allowed_tools field to restrict which tools a model may call.²³

Version 0.4.0 and deployment

Version 0.4.0, released January 28, 2026, rearchitected the software to separate the graphical interface from the core. It introduced llmster, the core packaged as a standalone server-native daemon that can be deployed on cloud servers or in continuous integration without a GUI, along with parallel requests to the same model using continuous batching instead of queueing, a new stateful /v1/chat REST endpoint, and a refreshed interface with chat export and split view.²⁴

Privacy

The desktop privacy policy states that when models are run locally, messages, chat histories, and documents are never transmitted from the user's system. The company receives data only when the app checks for updates, when the user searches for or downloads a model (in anonymized form), when the user emails support, and, if purchased and used, when requests are sent to paid cloud features such as cloud models or web search, which are processed transiently under zero-data-retention terms and not retained after the request completes.⁹ The policy also states that the app contains no telemetry or user-specific tracking, so the company cannot fulfill data-subject requests for copies or deletion of user data.⁹

Recent developments

Several changes in 2026 extended the product beyond the desktop. In April 2026 the Locally AI apps joined LM Studio to focus on Apple platforms, and in June 2026 LM Link, which lets a local machine use a model running on a remote device as if it were loaded locally, became available on iPhone and iPad through Locally, LM Studio's mobile app.¹⁷^•²⁵ In March 2026 LM Studio added support for NVIDIA DGX Station GB300, bringing open-model inference to a Blackwell-based workstation form factor outside the data center.¹⁷ In July 2026 the company announced LM Studio Bionic, an AI agent application built for open models.²⁶ The MLX engine continued to improve for agentic use: mlx-engine v1.8.5 added KV cache checkpointing, which the company reports dramatically improves performance for repeated, long-context agentic workflows.¹⁷

References

Topic: Encyclopedia › Technology and the built world › Computing and digital systems › Modern AI: foundation models, generative AI, and the AI industry › AI companies, people, and products › AI products and assistants › AI agents and developer tools

Initially written Sep 17, 2026 · Reviewed: Sep 19, 2026; Oct 15, 2026 · Edited: Oct 15, 2026 · Last review: Oct 15, 2026

Notice something wrong?