In reply to @mixedbreadai

Scaling laws are our greatest friend when running experiments: we can test ideas out to figure out which ones are worth scaling. But for ColBERT models, we didn't know which one to use! In 2025, there still were no flash-attention, long-context small ColBERT available!