posted an update —

8B as the reader

Apertus 1.5 8B is the reader. A small decision head sits on its hidden states and has to return a closed label, not an open answer.

The 70B does not fit a 24 GB GPU, so it is out. A full fine-tune of the 8B does not fit either. The base stays frozen, or takes a small adapter.

The sheaf model stays a separate line of work. It is not this prototype.