Base models trained on 1T high-quality tokens, demonstrating strong competitiveness among existing SOTA small models (<2B).