Inkling-Mini's surprise: a 3B-param model that genuinely needs all 48 effective layers

teortaxesTex · x · 2026-11-22

teortaxesTex finds Inkling-Mini "incredibly" interesting: he naively assumed the 3B-param model could be accelerated with adaptive looping, but analysis shows no redundancy was left on the table. Unlike typical models at this scale, it fails after just 1-2 passes and genuinely requires all 48 effective layers.

Original post →

More from Research

Research channel →