Mimir: A Sub-2B Hierarchical Reasoning Model Trained on Permissible Data Punches Above Its Weight Class
This work introduces Mimir, a 1.78B-parameter language model based on the Hierarchical Reasoning Model Text (HRM-Text) architecture, a recurrent-depth transformer whose layers are split into a low-level and high-level stack that are applied repeatedly in an interleaved schedule.
Peter Schneider-Kamp, Jacob Nielsen, Gianluca Barmina et al.
· 1 citation