When you don't want yo...
When you don't want your conversation history to stay on cloud servers
Have you ever had this experience: you chat with an AI about work-related secrets or personal privacy topics, and even though you know "the data won't be used for training," there's always that nagging feeling in the back of your mind – those conversation records are, after all, sitting on someone else's servers.
Then I noticed a friend of mine who works in information security uses AI more frequently than anyone else, yet I never see him open ChatGPT in a browser. When I asked, he sent me a link: "Install this – it runs completely locally and works even without an internet connection."
That link was Jan.
What is it?
Jan is an open-source, cross-platform AI desktop client that supports Windows, macOS, and Linux. Its core positioning is very clear: run large language models 100% offline, with all conversation data stored exclusively on your own device.
You can think of it as a "ChatGPT that lives inside your own computer" – no account registration required, no internet connection needed, and no data uploaded to the cloud. Open the app, pick a model, and start a conversation. It's that simple.
How is it different fr...
How is it different from ChatGPT?
The most fundamental difference is: your data belongs to you.
When you use ChatGPT, every conversation you have is sent to OpenAI's servers. With Jan, all models run locally – your conversation history, uploaded files, and all interactive data exist solely on your own hard drive.
Another difference: it doesn't restrict you to a single model. Jan supports downloading and running a wide variety of open-source large language models from Hugging Face, including mainstream models like Llama, Gemma, and Qwen. Depending on your hardware configuration and needs, you can choose models of varying parameter sizes – from small models of a few billion parameters up to larger models of several tens of billions.
Its technical architecture
Jan is powered under the hood by Cortex – an embeddable local AI inference engine that can run on any hardware. It has built-in support for multiple inference engines, including llama.cpp and TensorRT-LLM. This means you don't need to configure a Python environment, deal with CUDA dependencies, or wrestle with model formats – Jan encapsulates all of that for you.
It also comes with an OpenAI-compatible API server, allowing developers to call local models via API and integrate Jan's capabilities into their own workflows.
Its core features
-
Full conversation inte...
Full conversation interface: a ready-to-use chat UI with support for multi-turn conversations and conversation history management.
-
Model browser: browse and download open-source models from Hugging Face directly within the application.
-
Assistant framework: support for creating custom AI assistants optimised for specific tasks.
-
100% offline operation: all inference is performed locally, with no network connection required.
-
Cross-platform support: a consistent experience across Windows, macOS, and Linux.
-
Open source and free: licensed under Apache 2.0, with fully public source code.
Who is it for?Jan's po...
Who is it for?
Jan's positioning is clear:
-
If you have strict requirements for data privacy – such as handling business secrets, medical information, or legal documents – Jan is probably one of the most reassuring AI tools available. All data stays local, and nothing is uploaded to the cloud.
-
If you frequently work in offline environments – on planes, on the subway, in remote areas – Jan can run completely offline.
-
If you don't want to be tied down by API call fees – using Jan does not incur any API costs; you just need enough local computing power to run the models.
-
If you're a proponent of open-source software – Jan is licensed under Apache 2.0, with fully open source code that you can freely review, modify, and even build upon.
Its limitations
Of course, Jan has its boundaries. It requires local computing power – running large language models demands certain hardware specifications, especially for larger models that need sufficient RAM and VRAM.
Model selection requir...
Model selection requires some judgement on your part – unlike ChatGPT's "open and use" experience, you need to choose the right model based on your hardware configuration. However, the built-in model browser already lowers that barrier significantly.
Chinese language support depends on the selected model – Jan itself supports a Chinese interface, but the quality of conversation depends on how well the model you download handles Chinese. Choosing domestic models like Qwen can deliver a better Chinese-language experience.
A few final words
In an era where "all data is uploaded to the cloud," Jan takes the opposite path: it puts the power of AI back under your own control. You don't need to trust any cloud provider, worry about data breaches, or pay for API calls. All you need is a computer and a model willing to run locally.
As its developers say: "Making AI accessible to everyone, one commit at a time."
If you're also looking for an AI assistant that truly "belongs to you," why not check out Jan's official website or GitHub repository? After all, the best privacy protection is when your data never leaves your device in the first place.