Open source alternatives to LM Studio

What LM Studio does

LM Studio is a desktop app for macOS, Windows and Linux that finds and downloads open-weight models from Hugging Face and runs them on your computer with llama.cpp or, on Apple Silicon Macs, Apple’s MLX. It has a chat interface, offline chat with attached documents, support for MCP servers, and a local API with OpenAI-compatible endpoints that other apps can call, on the same machine or across the network. The lms CLI and a headless mode, llmster, run it without the GUI. A separate, newer app, LM Studio Bionic, is an agent that uses local models and can optionally use paid, US-hosted cloud models. The app is free to use; its JavaScript and Python SDKs and the lms CLI are MIT-licensed, but the app itself is not open source.

On privacy, LM Studio states that when you run models locally, messages, chat history and documents stay on your device; the app contacts its servers only for update checks, model search and downloads, and any cloud features you choose to use.

Why people look for an alternative

• Proprietary license. The terms grant a license for personal or internal business use only. You may not modify the app, build derivative works, reverse engineer it, redistribute it, or offer it to others as a hosted service.
• Features under the vendor’s control. The terms allow the company to add or remove features, change limits, or suspend or discontinue the software and its services.
• Serving models to others. llmster and the network API make it usable on a server for your own organization, but the license rules out running it as a service for third parties.
• Hardware and platform limits. On macOS only Apple Silicon is supported (macOS 14 or newer); Intel Macs are not. Windows on x64 needs a CPU with AVX2, and the Linux build is an AppImage that requires Ubuntu 20.04 or newer.

Open source runners keep prompts on your hardware in the same way. Whatever you choose, local models need enough RAM or GPU memory for the model size you pick, and the open-weight models that run on typical hardware are generally less capable than the top hosted models.

What to compare in an alternative

1. Desktop app or engine. A complete app with a chat window and model browser, or a model server that other chat clients connect to.
2. Hardware backends. Support for your GPU (CUDA, ROCm, Metal, Vulkan) or CPU-only inference, as documented by each project.
3. Model formats. GGUF through llama.cpp, MLX on Apple Silicon, or other formats and engines.
4. API compatibility. An OpenAI-compatible endpoint so existing tools and chat apps can use it.
5. Platforms and license. Linux servers, Docker, desktops or phones, and whether the whole stack is under an OSI-approved license.

Which one to choose

For a desktop app closest to LM Studio, Jan offers a chat interface, model downloads, local inference and an OpenAI-compatible server. Ollama downloads and serves models through a simple CLI, a desktop app and a local API that many chat apps can connect to. llama.cpp is the engine underneath Ollama, Jan, LocalAI and TextGen; its llama-server adds a built-in web chat and OpenAI-compatible endpoints. LocalAI is a self-hosted server for text, speech, image and video models behind OpenAI-compatible APIs. TextGen, formerly text-generation-webui, is a local app with several backends, LoRA fine-tuning and image generation for people who want detailed control. On phones, PocketPal AI downloads small models and runs them offline on the device.