Jan (app)

Jan is an open-source desktop application that runs large language models locally on a user's own computer, positioned as a privacy-focused alternative to hosted chat assistants. Released under the Apache 2.0 license, it runs on Windows, macOS, and Linux, and it is built so that, by default, data stays on the device and internet access is optional. The target user is someone who wants a polished, familiar chat experience — the kind hosted assistants trained everyone to expect — without sending their conversations to a third party. That framing aligns squarely with a self-sovereign approach to computing: own the hardware, own the weights, own the transcript.

How it works

Jan lets a user browse, download, and run open-weight models — Llama, Gemma, Qwen, and other families — from model hubs, then chat with them through a clean desktop interface. Under the hood it manages the local runtime (built on the llama.cpp engine family) so the user never assembles a serving stack by hand: no Python environments, no driver archaeology, no command line. Models arrive as quantized GGUF builds sized to fit consumer hardware, and the app surfaces which variants suit the machine's memory. It supports creating custom assistants with their own instructions, and integrates the Model Context Protocol (MCP) for connecting tools — while keeping execution resident on the local machine.

Local API and optional cloud

Jan also runs a local OpenAI-compatible API server, so other applications on the same machine can treat the locally hosted model as if it were a remote service — a quiet but important feature that turns a chat app into infrastructure. While the core promise is offline operation, Jan can optionally connect to external providers when the user explicitly chooses, presenting local and remote models in one interface. That design leaves the privacy trade-off where it belongs: a deliberate per-conversation decision in the user's hands, rather than a default of remote processing the user never consented to think about.

Where it fits

Hardware expectations deserve honest framing, because packaged apps can make local AI look more magical than physics allows. Small quantized models run acceptably on ordinary laptops with 8–16 GB of RAM; mid-size models want a discrete GPU with meaningful VRAM or an Apple-silicon machine with generous unified memory; and the largest open models remain out of desktop reach regardless of interface polish. Jan surfaces recommendations, but the underlying rule is universal: the model must fit in memory, and generation speed tracks the hardware, not the app. It is also worth understanding what an aggregator app implies — Jan curates and updates the runtime and model catalog, so users inherit its release cadence, a lighter but real dependency compared with assembling a stack by hand. For most people extending sovereignty from money to computation, that trade is sensible: the custody instinct that says verify your own node does not require compiling your own inference engine — it requires knowing where your data goes, and with Jan the default answer is nowhere.

Among local-AI tools, Jan occupies the packaged-app tier alongside LM Studio: maximum approachability, minimum configuration. Tinkerer front ends like text-generation-webui expose more engines and knobs; runtimes like Ollama serve the terminal-first crowd; server engines target multi-user throughput. Jan's differentiators within its tier are its open-source license and its local-first defaults. For the audience D-Central serves — people who already run their own node or miner and are extending that self-reliance to AI — an open, auditable chat app that works on a plain laptop is often the right first step; the heavier stacks can come later, if ever.

Compare local-AI apps in the sovereign self-hosting catalog.