Jan is a free, open-source app that works like a private ChatGPT on your own computer. You download a model once, and from then on you can chat offline, with your conversations stored on your machine.
It also connects to cloud models like GPT and Claude when you want more power, runs a local OpenAI-compatible API server, and supports projects, assistants and MCP tools.
This guide covers everything: install, picking a model that fits your PC, chatting, projects with files, cloud providers, the local API, settings and fixes for common errors. It is based on Jan’s official docs and changelog, up to Jan v0.8.8 (released 28 October 2026).
- What: a free, open-source (Apache 2.0) desktop app for Windows, Mac and Linux that runs AI models offline. No account needed.
- Install: download from jan.ai, the Microsoft Store or Flathub, then launch. Jan downloads its default model on first launch.
- More models: open Hub, check the “Fits / May be slow / Won’t fit” label, click Download. Paste a Hugging Face model ID to get any GGUF model.
- Cloud models: Settings > Model Providers, paste your OpenAI or Anthropic key.
- Local API: Settings > Local API Server > Start Server. Base URL
http://127.0.0.1:1337/v1. - Specs: about 8 GB RAM for 3B models, 16 GB for 7B, 32 GB for 13B (Jan’s Mac guidance).
- Slow? Turn the GPU on in Settings > Hardware, pick a smaller model or lower the context size.
What Is Jan AI?
Jan is an open-source ChatGPT replacement built by the Jan team (janhq on GitHub). It is built on llama.cpp and Tauri and licensed under Apache 2.0, so it is free for personal and commercial use.
- Local models: download and run models like Llama, Gemma, Qwen and gpt-oss from Hugging Face.
- Cloud models: connect OpenAI, Anthropic, Mistral, Groq, MiniMax and other providers with your own API key.
- Assistants and projects: save custom instructions and group chats with shared files.
- Local API: an OpenAI-compatible server at
localhost:1337for other apps. - MCP: connect tools and data sources through the Model Context Protocol.
- Privacy: with local models, chats and files stay on your computer.
Recent updates
| Version | Date | What changed |
|---|---|---|
| v0.8.8 | 28 Oct 2026 | Security release; local API checks request origins; bundled llama.cpp engine 0.7.0; fix for non-English Windows user names |
| v0.8.7 | 20 Oct 2026 | Switch to turn off the shell tool in chat; built-in subagents for the Jan agent; hosted Exa MCP default removed |
| v0.8.6 | 9 Oct 2026 | CUDA restored on Turing GPUs (GTX 16xx, RTX 20xx) with older drivers; MCP tool images reach the model |
| v0.8.5 | 8 Oct 2026 | llama.cpp engine ships inside the app; jan CLI no longer bundled; Trusted Hosts enforced on 0.0.0.0 |
Jan vs LM Studio vs Ollama
| Feature | Jan | LM Studio | Ollama |
|---|---|---|---|
| Open source | Yes (Apache 2.0) | No (free app) | Yes (MIT) |
| Chat app | Yes | Yes | Yes (desktop app) plus CLI |
| Cloud models in same app | Yes, many providers | Mainly local | Ollama cloud models |
| Local API port | 1337 | 1234 | 11434 |
| Projects with files | Yes | Chat with documents | Via other front-ends |
| MCP | Yes | Yes | Via apps |
| Best for | Open-source fans, mixing local and cloud | Model browsing and tuning | Developers and servers |
If you want to compare in depth, see our LM Studio beginner guide and our complete Ollama guide.
System Requirements
| System | Minimum | Notes |
|---|---|---|
| macOS | 13.6 or newer | 8 GB RAM for 3B models, 16 GB for 7B, 32 GB for 13B; Apple Silicon uses Metal automatically |
| Windows | Windows 10 or newer | GPU support for NVIDIA, AMD and Intel Arc |
| Linux | Most distributions | deb, AppImage and Flathub builds; GPU acceleration available |
Jan uses CUDA on NVIDIA cards and Vulkan for AMD, Intel and GPUs that cannot run CUDA 13. Not sure what your GPU can handle? Our VRAM guide for AI has a simple size chart.
Step 1: Install Jan
- Download Jan from jan.ai, or install it from the Microsoft Store (Windows) or Flathub (Linux). GitHub Releases has every build.
- Run the installer. On Windows run the .exe, on Mac open the .dmg and drag Jan to Applications, on Linux use the .deb or AppImage.
- Launch Jan. No account or sign-in is needed.
- Wait for the default model. On first launch Jan downloads its own foundation model, so you can chat straight away.
- Say hello. Type a message in the chat box to check everything works.
Step 2: Download the Right Model
- Open Hub from the left sidebar.
- Click a model to see its details. Jan shows a fit label for your hardware: Fits, May be slow or Won’t fit.
- Pick one that says Fits and click Download. Progress shows in the top bar, and you can cancel with the × button.
- Select it from the model selector in any chat.
Get any Hugging Face model
- Copy the model ID on Hugging Face, for example
owner/model-name-GGUF. Jan wants the ID, not the full URL. - Paste it into the Search bar on the Hub page.
- Choose a quantised file (Q4_K_M is a good default) and download.
- Gated model? Add a Hugging Face token under Settings > General > Others > HuggingFace Token.
You can also click Use this model on a Hugging Face model page and pick Jan under Local Apps. Our Hugging Face download guide explains which file to pick. Already have GGUF files? Import them through the Llama.cpp engine settings (or MLX on Apple Silicon).
| Your PC | Model size to start with |
|---|---|
| 8 GB RAM, no GPU | 1B to 3B |
| 16 GB RAM or 8 GB VRAM | 7B to 9B (Q4) |
| 32 GB RAM or 12 to 16 GB VRAM | 12B to 14B |
| 24 GB+ VRAM or 64 GB Mac | 27B to 32B |
Step 3: Chat, Assistants and Settings
- New chat: start a thread, choose a model from the selector and type.
- Assistants: save a name and instructions, such as “You are a friendly English tutor”, and reuse them.
- Context size: if Jan shows an “out of context” banner, click Increase Context Size or raise Context Size in model settings. More context uses more memory.
- GPU: under Settings > Hardware > GPUs, make sure the GPU switch is on.
- Web search and file upload: available as separate features in recent versions; see the Jan docs for the providers they use.
Step 4: Projects and Files
- Create a project: click New Projects in the left sidebar (or press ⌘ P on Mac) and give it a clear name.
- Assign an assistant so its instructions apply to every chat in the project. Click Edit to change it later.
- Add files: click Add in the Files panel. PDFs, Markdown, DOCX, XLSX, PPTX and code files work.
- Ask questions. Files are split into chunks and used as context in all chats in that project. The panel shows each file’s size and chunk count.
This is great for a study folder, a work handbook or a codebase you want to ask about privately.
Step 5: Add Cloud Models (Optional)
- Go to Settings > Model Providers.
- Pick a provider such as OpenAI or Anthropic and paste your API key.
- Choose that provider’s model from the selector and chat.
- Not listed? Use a custom endpoint for any OpenAI-compatible service.
Cloud models send your messages to that provider and bill your API account. Local models stay free and private.
Step 6: Use Jan’s Local API Server
- Download a local model first (llama.cpp, or MLX on Apple Silicon).
- Open Settings > Local API Server and pick it under Default Model Local API Server.
- Optional: set an API key under Configuration > API Key. Leave it empty to turn off auth.
- Click Start Server and wait for Server Status to show running.
- Point your app at
http://127.0.0.1:1337/v1.
Test it with curl:
curl http://127.0.0.1:1337/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{"model": "YOUR_MODEL_ID", "messages": [{"role": "user", "content": "Tell me a joke."}]}'
By default the server listens on 127.0.0.1, so only your PC can reach it. If you open it to your network, fill in Trusted Hosts and set an API key.
MCP Tools and the Jan Agent
Jan supports the Model Context Protocol, so a model can use tools such as file access or web search through MCP servers. MCP left experimental status in v0.6.9, and v0.8.7 added built-in subagents to the Jan agent.
Tool use works best with models that support tool calling. Since v0.8.7 you can switch off the shell tool in chat, which is a good idea unless you need it.
Where Jan Stores Your Data
| System | Default data folder |
|---|---|
| Windows | %APPDATA%/Jan/data |
| macOS | ~/Library/Application Support/Jan/data |
| Linux | ~/.local/share/Jan/data |
Models are big, so you may want them on another drive. Change it in Settings > General > Jan Data Folder.
Troubleshooting
| Problem | Fix |
|---|---|
| Very slow, GPU not used | Check nvidia-smi works, turn the GPU switch on in Settings > Hardware > GPUs, restart Jan |
ggml_vulkan error (AMD/Intel) |
Update your GPU driver |
ggml_metal error (Mac) |
Model too big for unified memory; use a smaller one |
error loading model |
Partial download; delete and download again |
| Out of memory | Smaller or more quantised model, lower context size, lower ngl |
| “Failed to fetch” / “Something’s Amiss” | Use models under about 80% of free RAM; lower ngl in Engine Settings |
| Jan won’t start | Quit stale Jan processes; rename store.json (chats and models are kept) |
| Port 1337 in use | Change Server Port in Settings > Local API Server > Configuration |
| Cloud 401 / 429 / 404 | Wrong key / quota hit / base URL missing /v1 |
jan command not found |
CLI is not bundled since v0.8.5; remove old PATH entry |
Logs live in logs/app.log inside the data folder, or open System Monitor in the footer and choose App Log. Jan keeps logs for 24 hours.
My Take
Jan is the best pick if you want a fully open-source ChatGPT-style app that mixes local and cloud models in one place. The fit labels in Hub make choosing a model easy for beginners, and projects with files are genuinely useful.
Power users who love tweaking every model setting may still prefer LM Studio, and developers often stick with Ollama. But for private everyday chat with an open codebase you can trust, Jan is a great default.
FAQ
Is Jan AI free?
Yes. Jan is free and open source under Apache 2.0. You only pay if you connect a paid cloud API.
Does Jan work offline?
Yes. Once a local model is downloaded, chats run fully offline on your computer.
Is Jan AI safe and private?
With local models, your chats and files stay in Jan’s data folder on your machine. Cloud models send messages to that provider.
Which model should I use in Jan?
Pick one that Hub labels “Fits”. On 16 GB RAM or 8 GB VRAM, a 7B to 9B model at Q4 is a good start.
What port does Jan’s API use?
Port 1337, with the base URL http://127.0.0.1:1337/v1.
Can Jan use my AMD or Intel GPU?
Yes, through Vulkan, with up-to-date drivers. NVIDIA uses CUDA, and Apple Silicon uses Metal.
Is Jan better than LM Studio?
Jan is open source and handles cloud providers well. LM Studio has deeper model tuning. Both are good free choices.




