Hertz-lm: > 6.6B parameters, 32-layer decoder-only transformer
> Context of 2048 input tokens (~4.5 mins)
> Predicts 15-bit compressed versions of hertz-codec tokens

→ View original post on X — @reach_vb