Tool
Jan
An open-source desktop chat app that downloads and runs open-weight models on your machine (llama.cpp under the hood), with optional hooks to cloud providers when you want them — ChatGPT-shaped UI, local-first by default.
What it is
Jan (janhq / Menlo Research) is a cross-platform desktop client: browse or pull models, chat locally, expose an OpenAI-compatible local API, and optionally connect cloud models (OpenAI, Anthropic, and others). Inference leans on a bundled llama.cpp-style path rather than asking you to assemble a server stack first. Downloads live at jan.ai and GitHub Releases.
What it replaces
It replaces a hosted ChatGPT/Claude website for a lot of everyday drafting when open weights are good enough and you want prompts on-disk. It also replaces piecing together Ollama + a separate UI if you want one installer-shaped desktop experience. It does not replace a frontier API when you need the strongest reasoning, and it does not replace a properly sized inference server for a team.
License
The janhq/jan LICENSE is based on the Apache License, Version 2.0, with an added request for attribution in user-facing documentation and materials. GitHub’s license metadata may show “Other” because of that clause — it is not unmodified Apache-2.0 text. The app license still does not cover the model weights you download; each model card has its own terms.
Where it falls over
- VRAM and RAM decide whether “local” feels fast or theatrical.
- Optional cloud connectors send prompts off-box the moment you enable them — local-first is a setting, not a guarantee.
- Desktop apps still update; pin versions if you need a frozen compliance story.
- Model licenses (Llama, Gemma, Qwen, etc.) can restrict commercial use independently of Apache on the app.
- A polished UI is not a substitute for evaluating output quality on your tasks.
Who it's for
People who want a desktop ChatGPT-shaped experience on their own hardware without assembling Ollama + Open WebUI first. Fine for private drafting and local API experiments. Wrong tool if you need multi-tenant GPU serving (see vLLM) or a zero-ops frontier chat subscription.
Hardware that usually matters
The app is free to download. The recurring cost is the machine: VRAM first, then system RAM, then disk for weights. Example Amazon listings — not endorsements; stock and price move.
- GIGABYTE GeForce RTX 3060 Eagle OC 12GB — entry 12 GB VRAM; the pattern many beginners use for 7B–13B quantized models.
- CORSAIR Vengeance LPX DDR4 32GB (2×16GB) — cheaper DDR4 system RAM path when layers spill or you CPU-offload on older boxes.
- Samsung 990 PRO 2TB NVMe SSD — model files eat disk; a fast NVMe keeps pulls from feeling stuck.
Some links are Amazon Associates. We may earn a commission if you buy.
Related in this cluster
- Ollama — CLI/API-first local runner many UIs sit on
- LM Studio — another desktop local-model browser (proprietary app)
- llama.cpp — the MIT C/C++ engine Jan’s local path sits near
- Open WebUI — self-hosted browser UI instead of a desktop app
Verdict
Use Jan when you want a desktop, local-first chat client and Apache-shaped terms on the app are enough. Read model licenses separately, treat cloud connectors as intentional off-box, and size VRAM honestly. Stars are not a substitute for running your prompts on your hardware.