Jan: the open-source alternative to ChatGPT

#Jan in two words: what it is in 2026

Jan is a free desktop application developed by Menlo Research that replaces ChatGPT with models running on your machine. Its code is released under the Apache 2.0 license on GitHub. It runs on Windows, macOS (Apple Silicon only), and Linux; downloads GGUF-format models from a Hub connected to Hugging Face; runs them with llama.cpp; and exposes an OpenAI-compatible API on port 1337. It can also connect to cloud models (OpenAI, Anthropic, Google, Groq, Mistral, OpenRouter) in the same interface, connect MCP servers, and perform web searches. For everyday use, no account or command line is required. This guide installs Jan, helps you choose a model suited to your memory, and shows you where its strengths end.

#Jan, LM Studio, or Ollama: what sets them apart

All three tools run the same GGUF models; they differ in their interface, license, and typical use. The table is limited to elements verifiable in their official documentation.

The choice depends on how you use it. Jan works if you want a complete interface, open-source code, and the ability to mix local and cloud models without changing tools. If you automate, script, or deploy on a server, Ollama or llama-server remain more natural choices. The comparison guide details these trade-offs.

#Install Jan on Windows, macOS, or Linux

Installation does not require an account. Here are the official prerequisites to check before downloading: they determine which model you can load afterward.

1. 01
   Download the installer

   Go to jan.ai or the GitHub releases page. Windows: .exe file. macOS: .dmg to drag into Applications. Linux: Flatpak preferred, otherwise .deb or AppImage.
2. 02
   Launch Jan for the first time

   No account creation. On first launch, Jan downloads its default model; once the download is complete, chat is ready to use. The first launch also lets you choose your usage measurement preferences.
3. 03
   Open the Hub for more models

   The Hub is accessible from the sidebar. Search for a model, click Download, then select it in the model selector of any chat.

flatpak install flathub ai.jan.Jan
flatpak run ai.jan.Jan

#Choose a model that fits in your memory

The Hub displays a fit indicator for each model: Fits, May be slow, or Won't fit, calculated from your hardware without downloading anything. Quantizations are grouped into Small, Balanced, and Large, with a Recommended label for the default download. Check these indicators before considering any other criteria.

To give you a sense of scale, here are the site's approximate Q4 sizes (weights only, before context) and the figures Jan provides for Macs; the table cross-references both.

The in-house Jan-v3-4B model, released under the Apache 2.0 license, is a reasonable starting point: 4 billion parameters, a native context window of 262,144 tokens, and, according to Jan, at least 8 GB of RAM in Q4 quantization (16 GB recommended in Q8). It is a lightweight model: for summarization or long-form writing, a model with 7 to 14 billion parameters will do better if your memory allows it.

#First chat, assistants, and projects

1. 01
   Open a new chat

   Select the model from the selector at the top of the chat. A local model is loaded into memory the first time you use it; allow more time for a large model.
2. 02
   Write the message

   Send your message; the response appears as it is generated. The model parameters (temperature, context) are documented in Jan's “Model Parameters” documentation.
3. 03
   Attach a file

   The plus button lets you add documents, images (a vision-enabled model required), or WAV and MP3 audio files (an audio model required). With a local model, the files stay on your machine.

Two features change everyday use. Assistants are reusable sets of instructions: the equivalent of a saved system prompt that you activate from the input field. Projects group conversations around shared instructions and files: a document added to a project is split into chunks and indexed, then made available in every conversation in the project. It's lightweight RAG, sufficient for querying a few PDFs without building a pipeline. See the guide to system prompts for advice on writing an assistant's instructions.

#Connect a cloud model (and what that entails)

Jan lets you mix local and cloud models in the same interface: useful for comparing a local response with one from a large model, or for keeping a single chat client. Open Settings, then Model Providers, choose a provider (OpenAI, Anthropic, Google, Groq, Mistral, OpenRouter, Azure OpenAI, Hugging Face), and paste in your API key. Any OpenAI- or Anthropic-compatible endpoint can also be added as a custom endpoint. The cloud model then appears in the selector.

#Use Jan as a local API server

Jan includes an OpenAI-compatible server powered by llama.cpp. Open Settings, then Local API Server, and click Start Server. It is ready when the logs show it listening on http://127.0.0.1:1337. Your scripts, code editors, or OpenAI-compatible applications can then use it by changing the base URL.

curl http://127.0.0.1:1337/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer secret-key-123" \
  -d '{"model": "YOUR_MODEL_ID", "messages": [{"role": "user", "content": "Bonjour"}]}'

Replace YOUR_MODEL_ID with the ID of a model available in your Jan, and the key with the one you defined. Configuration options deserve attention, especially those related to networking.

#Common local server errors

Connection refused

The server is not running, or your application is targeting the wrong host or port.

401 Unauthorized

The API key is missing from or incorrect in the Authorization header.

404 Not Found

The model identifier does not match any model available in Jan, or the API prefix is misspelled.

CORS error in the browser

Make sure the CORS option is enabled in the server settings.

#llama.cpp or MLX on Mac

On Apple Silicon, Jan offers two engines. llama.cpp is the stable engine, available on all platforms in GGUF format. MLX is an engine for Apple chips (macOS 14 minimum) that uses Metal, but Jan labels it experimental: no embeddings, so no RAG; reasoning output is unsupported; and some recent architectures will not load. The documentation recommends starting with llama.cpp and trying MLX for a compatible model if you want speed.

#Privacy: what stays with you

Jan says it collects no data before you consent: you choose your usage-measurement preferences on first launch and can change them later in the settings. If you agree, the publisher counts usage metrics, such as active users, linked to a random identifier, through PostHog hosted in the European Union. It says it never reads your conversations or logs your prompts. Your data (models, conversation threads, settings, logs) is stored in a local folder: on macOS, ~/Library/Application Support/Jan/data; on Linux, ~/.local/share/Jan/data. This description comes from the publisher; for a regulated environment, verify your workstation’s outbound traffic.

#Is Jan right for you?

You want open-source code

The repository is licensed under Apache 2.0: you can read, fork, and redistribute the code in compliance with the license, with attribution requested.

You're mixing local and cloud

One interface for your local models and cloud providers, with shared assistants and projects.

You avoid the terminal

Installation, model downloads, and the local server are all managed from the interface.

You automate or deploy on a server

Prefer Ollama or llama-server, which are better suited to services that run without an interface.

You're on an Intel Mac

Jan is not supported: choose another tool.

• Jan: the open-source alternative to ChatGPT, 100% local (tutorial)
• Ollama vs. LM Studio vs. Jan vs. GPT4All
• LM Studio for beginners: your first local chat
• Master system prompts
• llama-server: local OpenAI API with llama.cpp

• Source: Jan Desktop’s official documentation
• Source: Jan’s GitHub repository
• Source: Jan's local API server
• Source: Jan's approach to privacy