Some nice quantization work makes our Gemma 3 model even more compact.

→ View original post on X — @jeffdean