We tried out Google’s new family of multi-modal models with variants compact enough to work on local devices. They work well.

Gemma 4 made local LLMs feel practical, private, and finally useful on everyday hardware.

I don't really know why, though.