This research introduces a novel training pipeline that creates small language models with competitive reasoning abilities while reducing computational demands and addressing privacy concerns.

• Efficient alternatives to large language models that can be deployed on edge devices
• Innovative training approach that enhances reasoning capabilities in smaller models
• Dual focus on both text-only and multimodal small language models
• Security benefits through reduced computational requirements and better privacy protection

This research matters for security professionals seeking AI solutions that offer powerful reasoning capabilities without the privacy and computational challenges of larger models.

InfiR: Crafting Effective Small Language Models and Multimodal Small Language Models in Reasoning