4. Jet-Nemotron A hybrid-architecture LM family: starting from a frozen full-attention model, the authors search for where to keep full attention, which linear-attention block to use, and which hyperparameters match hardware limits.

→ View original post on X — @dair_ai