
Ollama
Run open LLMs entirely on your own machine, offline.
Every prompt you send to a mainstream cloud chatbot can be logged, tied to your account, and fed back into training the next model. Private AI flips that: either the model runs locally on your own device so nothing leaves it or it is hosted by a company that encrypts your history and contractually refuses to train on you. The trade-off is real, though. Local models are less capable than frontier cloud systems and need decent hardware, while privacy-first hosted assistants ask you to trust a policy instead of your own machine.
Start by deciding between local and hosted. Running a model yourself, with a tool like Ollama or llamafile, is the strongest form of privacy because your prompts never touch a network but it leans on your hardware: several gigabytes of RAM and ideally a modern GPU for anything beyond small models. If that is too much, a privacy-first hosted assistant like Lumo, Duck.ai, or Euria keeps things easy while limiting or encrypting what the provider retains, at the cost of trusting a policy rather than math. Favour tools with open weights and open-source code so the privacy claims can be independently checked and match the model size to what your device or budget can actually handle.
Here are some things to look for.

Run open LLMs entirely on your own machine, offline.

Proton's private assistant that keeps no logs of your chats.

A whole LLM packed into one offline executable file.


Ente's offline chat app that runs models on your device.

Self-contained local model runner with a built-in UI.


Offline ChatGPT alternative that runs models on your device