Jan: An Open-Source Way to Run Models on Your Desktop
Jan is a desktop app that runs open-weight models locally, with the source open and the data staying in files you own. What it does, how it compares to the closed alternatives, and where its openness costs you something.
Key takeaways
- Jan is an open-source desktop application for running local models on Windows, macOS and Linux, with a bundled engine so there is nothing to install first.
- Its distinguishing argument is verifiable openness: the code can be read, and your conversations and settings live in plain local files rather than an opaque store.
- It can also expose a local API server, which turns it from a chat window into a backend other applications can call.
- It can connect to remote providers too — useful, and the setting to leave alone if the point is that nothing leaves the machine.
- Compared with closed alternatives it trades some polish and feature breadth for auditability. That is the whole decision.
What it is
Your private ChatGPT, free, on your own machine in an hour — LM Studio, Ollama, Open WebUI, your documents, no cloud.
- Lifetime online access
- PDF + files
- 30-day refund
Jan belongs to the same category as the other desktop apps: download, pick a model, chat, with the inference engine bundled so no separate service is required. Underneath it runs the same open-weight models in the same GGUF format as everything else in this space, which means performance is a function of your hardware, not of the app.
What differentiates it is the licensing and data posture. The codebase is open, and the artefacts it produces — threads, assistants, settings — are files on your disk. For anyone who has to explain to a compliance officer where the conversations physically are, that is a materially different answer from "in the application".
What you get
- A model library with one-click downloads, plus the ability to import a GGUF you already have.
- Assistants: saved system prompts and parameters for recurring tasks.
- A local API server exposing an OpenAI-compatible endpoint, so scripts and other tools can use the model Jan is running.
- Remote provider support, for people who want one window across local and cloud models.
- Extensions, which is where capability grows over time rather than being fixed at install.
The API server is the underrated feature. It means the desktop app you already trust can be the backend for a script, an editor plugin or an internal tool, without installing a second runtime. Point any OpenAI-compatible client at the local port and it works.
Jan, Msty and LM Studio
| Jan | Msty | LM Studio | |
|---|---|---|---|
| Source code | Open | Closed | Closed (engine open) |
| Bundled engine | Yes | Yes | Yes |
| Data on disk | Plain local files | Application store | Application store |
| Local API server | Yes | Yes | Yes |
| Document chat | Available | A headline feature | Limited |
| Model tinkering depth | Moderate | Moderate | The most granular |
| Pick it when | Openness is a requirement | You want the fastest start | You want to compare quants precisely |
If you are choosing between the two closed options rather than against Jan, LM Studio vs Ollama covers that ground.
What your machine needs
Identical to every tool in this category, because the constraint is the model, not the app: roughly 8 GB of memory for a 7B–8B model at 4-bit, 16 GB for 12B–14B, 32 GB and up for 24B–32B. A discrete GPU buys speed rather than capability. The numbers by model are in VRAM requirements, and the calculator answers it for a specific model.
The honest limits
- Feature breadth trails the closed apps. Openness has a development cost, and it shows in the periphery rather than in the core.
- Extensions vary in maturity. A capability that exists is not automatically a capability that is solid.
- Cloud providers are one click away. If the requirement is strict locality, do not configure them.
- It does not make hardware faster. Same model, same machine, same tokens per second as any other GGUF runner.
Verdict
Jan is the desktop app to choose when "open source" is a requirement rather than a preference: readable code, local files, a real API server, and no separate engine to install. Accept a slightly narrower feature set than the closed alternatives, keep remote providers unconfigured if locality is the point, and you have a local AI setup you can actually audit.
Frequently asked questions
Is Jan free and open source?
Yes, the application is open source and free to download. Check the specific licence in the repository if you plan to redistribute or embed it.
Does Jan need Ollama?
No. It bundles its own inference engine. It can also connect to other endpoints if you already run one.
Can Jan work offline?
Yes, once a model is downloaded. Model downloads, update checks and any remote provider are the only parts that use the network.
Where does Jan store my conversations?
In local files on your machine, which is part of its appeal: you can inspect, back up or delete them without going through the application.
Jan or LM Studio?
Jan when open source and auditable local data matter. LM Studio when you want the most granular control over models and quantisation, and accept closed source.
Can other apps use Jan's model?
Yes, through its local OpenAI-compatible API server. Point any compatible client at the local port and it works, without installing a second runtime.
A current option for local AI: GMKtec EVO-X2 64GB / 1TB (Ryzen AI Max+ 395). Match memory to your model and software. A mini PC is a complete PC alternative; Mac/MLX and CUDA instructions require compatible hardware.
Amazon Check GMKtec EVO-X2 64GB / 1TB (Ryzen AI Max+ 395) price →As an Amazon Associate, BestLLMfor earns from qualifying purchases, at no extra cost to you. It does not influence our independent rankings.
Found an error or have feedback? Let us know — it helps everyone who reads this guide.