Experiment calibrates 48 attention heads down to 12, swapping the rest for band-diagonal sparse attention

ostrisai · x · 2026-09-18

Original post →

More from Models

Models channel →