Jan is a desktop chat application for running large language models on your own machine, with the option to connect to cloud models as well. You download open models such as Llama, Gemma, Qwen, and GPT-oss from Hugging Face and run them locally, or add API access to hosted models from OpenAI, Anthropic, Mistral, Groq, MiniMax, and others, and switch between them in one interface.
Beyond chat, you can create custom assistants for specific tasks, connect tools through the Model Context Protocol (MCP), and expose the loaded models to other applications through an OpenAI-compatible local server. When only local models are used, everything stays on the device.
Local inference runs on a bundled llama.cpp engine, with builds for CPU, Vulkan, Metal, CUDA, and ROCm. The app is built with Tauri and is available for Windows, macOS, and Linux. Model size is bounded by memory: roughly 8 GB of RAM for 3B-parameter models, 16 GB for 7B, and 32 GB for 13B.
Features
- Local models: download and run open models such as Llama, Gemma, Qwen, and GPT-oss from Hugging Face
- Cloud models: connect to OpenAI, Anthropic, Mistral, Groq, MiniMax, and other hosted providers
- Custom assistants: create specialized assistants for particular tasks
- OpenAI-compatible API: a local server that lets other applications use the models loaded in Jan
- MCP support: Model Context Protocol integration for tool use and agentic workflows
- Local-only mode: everything runs on the device when only local models are used
- GPU acceleration: engine builds for Vulkan, Metal, CUDA, and ROCm alongside CPU
