0007
Ling 3.0 tiny · MLA
00 · Periodic full attention
Every fourth Ling block leaves recurrent KDA
Ling 3.0 tiny repeats a 3 KDA + 1 MLA rhythm. This scene follows the MLA block, where the token can directly attend over cached history while query and key/value information pass through compact low-rank bottlenecks.
Implementation
hidden_states → MLA attentionThe 3:1 layer rhythm and all dimensions shown here come from the released Ling checkpoint. Token values and glow intensities are explanatory.