[kristianp]

> The bigger size is probably from the bigger vocabulary in the tokenizer.

How does that affect anything? It still uses 16 bit floats in the model doesn't it?