<p>Oracle Bone Inscriptions (OBI), the earliest mature Chinese writing system, exhibit highly abstract pictographic structures and intricate semantic mappings, posing formidable challenges for automated recognition and interpretation. This study introduces OBIFlo-SAM, an efficient, lightweight, and robust cross-modal multi-task framework integrating the pre-trained Florence-2 model with the Segment Anything Model (SAM). Employing a weighted Focal Loss fine-tuning strategy, the framework achieves high-precision semantic generation, text alignment, object detection, and zero-shot segmentation, addressing low-resource scenarios, class imbalance, and complex backgrounds. The OracleVQA dataset, combining multimodal question-answer pairs with detection annotations, provides high-quality training data. Experimental results demonstrate that OBIFlo-SAM significantly outperforms CLIP-based and general-purpose large language models in semantic recognition while excelling in detection and segmentation. This work establishes an intelligent pathway for digital preservation and automated interpretation of ancient scripts, providing foundation for cultural heritage conservation and scholarly research.</p>