Jan vs LM Studio: Licence, Cloud Keys and Where Each Wins
September 29, 2026
Jan vs LM Studio looks like a coin toss. Both are desktop apps that download open models, chat with them offline, and run a local server other tools can call.
The real differences sit underneath, and as of September 2026 they come down to three things:
- Licence. Jan is open source under Apache 2.0. LM Studio is a closed-source app with MIT-licensed tooling around it.
- What "cloud" means. Jan lets you paste your own API keys for OpenAI, Anthropic, Gemini and others. LM Studio's cloud option is its own hosted service for open models.
- Maturity on a Mac. LM Studio treats MLX as a first-class engine. Jan ships MLX too, but its own docs call it experimental.
If you want the short version: pick Jan if you care about the source code or want local and frontier models in one window. Pick LM Studio if you're on Apple silicon and want the most polished model runtime.
Jan vs LM Studio: quick comparison
| Jan | LM Studio | |
|---|---|---|
| Licence | Apache 2.0, whole app | Closed-source app; CLI and SDKs are MIT |
| Maker | Menlo Research | Element Labs |
| Local engines | llama.cpp, plus MLX (experimental) on Mac | llama.cpp, plus MLX on Apple silicon |
| Cloud models | Your own keys: OpenAI, Anthropic, Gemini, Groq, Mistral, OpenRouter and more | LM Studio Secure Cloud (open models, Zero Data Retention), through Bionic |
| Local API | OpenAI-compatible, default port 1337 | OpenAI-compatible plus Anthropic-compatible, default port 1234 |
| MCP | Yes, Jan is an MCP host | Yes, since version 0.3.17 |
| Agent features | Cowork and Jan Agent, both in preview | Bionic, a separate agent app |
| Platforms | Apple silicon Mac, Windows x64, Linux | Apple silicon Mac, Windows x64 and ARM, Linux x64 and ARM64 |
The licence is the difference that won't change
Features converge. Licences don't.
Jan's repository carries an Apache 2.0 licence. You can read the code, audit what the local server does, fork it, and ship a modified build. Some comparison articles still list a different licence for Jan, so trust the LICENSE file over a listicle.
LM Studio's desktop app is proprietary. Its terms of service (version dated August 23, 2026) grant a licence to use the software "solely for Your personal and / or internal business purposes." The same terms forbid modifying the app, reverse engineering it, or offering it as a hosted service to third parties.
So internal business use is covered. Building a product on top of the app itself is not.
The open-source pieces of LM Studio are real, though. The lms CLI, the TypeScript and Python SDKs, and the MLX engine are all MIT-licensed on GitHub.
| Question | Jan | LM Studio |
|---|---|---|
| Can I read the app's source? | Yes | No. The CLI, SDKs and MLX engine only |
| Can I use it at work? | Yes, under Apache 2.0 | Yes, for internal business purposes per its terms |
| Can I fork and redistribute it? | Yes, with licence notices | No |
| Can I resell it as a hosted service? | The licence allows it | No, the terms forbid it |
Where Jan is stronger
One window for local and frontier models. Under Settings, you add an API key for a provider and its models show up next to your local ones. Jan's docs list Anthropic, Gemini, Groq, Mistral AI, OpenAI, Azure OpenAI, OpenRouter, Hugging Face, and custom OpenAI-compatible endpoints.
So you can draft with a small local model, then switch to a Claude or GPT model when the task gets hard. Jan's privacy page is plain about the trade-off: cloud models need to see your messages to work, and Jan "just makes the introduction."
A server you can configure in detail. Jan's local API server listens on 127.0.0.1:1337 by default. You can set an API key, bind it to your local network, restrict trusted hosts, and let MCP tool calls run on the server side.
Linux packaging. Jan recommends Flatpak on Linux, which updates through Flathub with the rest of your system. There are also .deb and AppImage builds.
Its own small models. Jan's team trains compact models (Jan-v1, Jan Nano, Jan-Code-4B and others) that it describes as built for agents and research on a personal computer.
| Need | Why Jan fits |
|---|---|
| Audit or modify the code | The whole app is Apache 2.0 |
| Mix local models with GPT, Claude or Gemini | Bring-your-own-key providers built in |
| Share a model on your home network | Host binding, API key and trusted hosts in settings |
| Run on Linux with system updates | Flatpak via Flathub |
Where LM Studio is stronger
MLX on a Mac. LM Studio runs both llama.cpp and Apple's MLX on Apple silicon, and its MLX engine is a maintained open-source project. Jan also added MLX, but its MLX docs flag it as experimental: no embeddings, reasoning output not yet wired, and some newer architectures fail to load. Jan's own advice is to start with llama.cpp if you're unsure.
If you bought a Mac to run local models, that can settle it. Our AI on Mac breakdown shows where MLX sits in the bigger picture.
Finding the right model. LM Studio's Discover tab searches Hugging Face directly. You can type a keyword, a user/model string, or paste a full Hugging Face URL. The docs also explain the quantisation labels in plain language, which saves you from downloading the wrong file.
A broader API surface. LM Studio's OpenAI-compatible endpoints include Responses, Chat Completions, Completions and Embeddings. Because it implements the Responses endpoint, tools like Codex can point at it directly. It also exposes Anthropic-compatible Messages endpoints, and it can run headless as a service.
LM Link. LM Link lets you run a model on your desktop and use it from your laptop, built with Tailscale and end-to-end encrypted. LM Studio's iPhone app, Locally, can connect to it too. Jan's closest option is binding its server to your local network.
Hardware coverage. LM Studio's system requirements cover Windows on ARM (Snapdragon X Elite) and Linux ARM64 as well as x64.
| Need | Why LM Studio fits |
|---|---|
| Local models on Apple silicon | MLX is a first-class engine, not a preview |
| Browsing and picking models | Built-in Hugging Face search with quantisation guidance |
| Pointing Codex or Anthropic-style clients at a local model | Responses and Anthropic-compatible endpoints |
| Using a desktop GPU from a laptop | LM Link across your own devices |
| Windows on ARM | Supported |
Both are becoming agent apps, at different speeds
Neither tool is just a chat window any more. The question is how far along each one is.
LM Studio's agent is Bionic, and it's a separate app from LM Studio. The docs describe it as an agent for coding, research and work with documents. It can use local models, models on another device through LM Link, or open models in LM Studio Secure Cloud. Its docs say you can keep using the original LM Studio app alongside Bionic for "advanced low-level configuration."
Cloud use in Bionic needs an account and credits. Local and LM Link models don't. The pricing page lists a free plan for local use and paid subscriptions for cloud inference.
Jan has two agent surfaces, and both carry a preview label. Cowork takes a multi-step task, reads and writes files in a folder you attach, runs commands, and shows you a diff of its changes. It ships only in Jan's nightly builds for now. Jan Agent brings the same idea to the terminal and shares providers, MCP servers and skills with the desktop app.
Both apps are Model Context Protocol hosts, and both document how to act as a local backend for Claude Code. If connecting tools is your main use, our guide to MCP clients covers what to check.
| Agent feature | Jan | LM Studio |
|---|---|---|
| Main agent product | Cowork in the desktop app, Jan Agent in the terminal | Bionic, a separate desktop app |
| Status | Preview, nightly channel | Available to download |
| Cloud option for the agent | Your own provider keys | LM Studio Secure Cloud, credit-based |
| Local backend for Claude Code | Yes, per-tier model assignment | Yes, via its integrations guide |
Hardware: check before you download
Both apps need an Apple silicon Mac. Jan's Mac install page and LM Studio's requirements page both say Intel-based Macs aren't supported.
| Jan | LM Studio | |
|---|---|---|
| macOS | 13.6 or newer, Apple silicon (MLX needs macOS 14) | 14.0 or newer, Apple silicon |
| Windows | Windows 10+, x64 with AVX2 | x64 with AVX2, or ARM |
| Linux | x64 via Flatpak, .deb or AppImage | x64 or ARM64 AppImage, Ubuntu 20.04+ |
| Memory guidance | 8GB for about 3B models, 16GB for about 7B, 32GB for about 13B | 16GB+ recommended; 8GB Macs should stick to smaller models |
Those are the vendors' own figures. Any tokens-per-second number you find online comes from one machine, one model and one build.
How to decide
Choose Jan if:
- You need open-source software, for policy reasons or because you want to audit it
- You pay for OpenAI, Anthropic or Gemini and want those models next to your local ones
- You're on Linux and want Flatpak updates
- You're happy to try preview agent features on a nightly build
Choose LM Studio if:
- You're on Apple silicon and want MLX without an "experimental" label
- You spend a lot of time comparing and downloading models
- You want to share one strong machine's models with your laptop
- You want an agent app that has shipped, with an optional cloud for big open models
Run both if you can't decide. Each manages its own model folder and serves on its own local port.
If you want a browser front end instead of a desktop app, Open WebUI and its alternatives are the next place to look. For the other half of the local-runtime question, see our LM Studio vs Ollama comparison.
Where Taku fits
Jan and LM Studio both answer "how do I run a model?" Most people get stuck one step later: what do I actually do with it?
Taku is an AI-native desktop workspace built for that step. You mirror an AI app, skill or workflow that someone already got working, run it, remix it for your own files, and keep it. It doesn't replace a model runtime like either of these. It's a different layer. Browse the free apps to see the shape of it. Taku is in Beta, and the Mac app is available now.
FAQ
Is Jan open source?
Yes. Jan's GitHub repository is licensed under Apache 2.0, and the project is maintained by Menlo Research. That covers the whole desktop app, not just a CLI or SDK.
Is LM Studio free for work use?
LM Studio's terms grant a licence for personal and internal business use. Its pricing page lists a free plan for running local models, with paid plans for cloud inference in Bionic. Check the current terms before rolling it out company-wide, since they're versioned and can change.
Which is better on a Mac, Jan or LM Studio?
On Apple silicon, LM Studio has the edge because MLX is a first-class engine there, while Jan labels its MLX support experimental. Both also run llama.cpp, and both need an Apple silicon Mac.
Can Jan use ChatGPT or Claude models?
Yes, with your own API key. Add the key under Model Providers in Jan's settings, and those models appear in the model picker. Your messages then go to that provider under its terms.
Can I use Jan or LM Studio as a backend for other apps?
Yes. Jan serves an OpenAI-compatible API on port 1337 by default. LM Studio serves OpenAI-compatible and Anthropic-compatible endpoints on port 1234. Point your tool's base URL at either one.