I trained a vision-language model that answers typed questions about an image. It uses also choice, score and noul, like Jev.

I thought, why Jev only processes text? There should be the possibility to process an image too. I will experiment with some pictures and questions in the next days to see, how well it performs in real life.

On my M1 Pro a request with six questionut 400 ms p95, about 60 ms on a desktop GPU. The server encodes the image once and scores each option as a short suffix against the KV

Looking forward to answer your questions! :)