I got tired of sending every text I translate to Google/DeepL. Even with all the opt-out options and privacy policies, it never felt right especially for some work documents, personal writing, or anything sensitive. So I decided to build this tool, which lets me use LLMs for context translations and also a standard translation engine like Argos. It works with Ollama, llama.cpp or argos-translate, and you can configure the model you want to use.

loqi translate --from it --to en "Ciao mondo"

Obviously, the quality of the translation depends entirely on the model used, but I've noticed that you can get good, if not excellent, results even with a small model (such as Gemma 4 E2B or Phi4-min).

So there you have it: Loqi is open source, cross-platform (MacOS, GNU/Linux, Windows), written in Go with Bubble Tea for the TUI. It allows the model to translate individual sentences or process entire files (whether plain text, Markdown or JSON).

I used LLMs to help write parts of the code, including self-hostable ones (those that run on AMD GPU with 16 GB of VRAM), but I've tried to set up the project as much as possible, and there's probably a lot more work to be done and some ideas to implement.

I'd be more than happy to accept contributions.