We propose RepBlend, a multimodal dataset distillation framework that mitigates modality collapse via representation blending and symmetric projection matching, achieving superior performance and generalization.