Very good, it now works for me too.

Indeed, special characters and German umlauts are transmitted webcoded, but apparently this does not bother the AI very much.
Thanks idle, for the libs and your effort!
The calculation with the model I use "ggml-model-gpt4all-falcon-q4_0" (about 4 GB) runs amazingly fast. i don't have to wait much longer than with chatgptx on the web.
It's a pity that today is the day when I have to prepare my tax return, otherwise I would have dealt with it more intensively right away. I would first see if one can do it without the web server.