Apriel-H1: Towards Efficient Enterprise Reasoning Models (arxiv.org) 1 points by guiriduro 7mo ago ↗ HN
[–] guiriduro 7mo ago ↗ Apriel-H1-15b-Thinker-SFT uses incremental distillation from Apriel-Nemotron-15B-Thinker, selectively replacing less critical attention layers with linear Mamba blocks to reduce computational complexity while preserving reasoning quality.
1 comment
[ 2.2 ms ] story [ 13.3 ms ] thread