We approach the task of beta-amyloid detection in Alzheimer’s patients and propose a multimodal contrastive and distillation-based learning framework integrating PET (using AV45 and PIB tracers) and 3T MRI data from 790 participants in the OASIS-3 cohort. Built on BiomedCLIP with efficient LoRA fine-tuning, our model leverages cross-modal and self-attention mechanisms with soft triplet loss and adaptive margin to align PET–MRI embeddings. Through knowledge distillation, we transfer PET-guided representations into MRI-only embeddings. A lightweight MLP predicts amyloid positivity from PET-guided MRI representations and MRI embeddings. Our preliminary results demonstrate that PET-guided MRI successfully transfers knowledge to an MRI-only model, showing substantial relative improvement (+8.3% F1 score) compared to models trained without distilled knowledge.