This paper investigates the use of multimodal large language models (LLMs) for classifying social media posts in disaster response. Using the CrisisMMD dataset, we evaluate two proprietary models—GPT-4o and GPT-4o mini under zero-shot and few-shot setting, and evaluate the open-source LLaMA 3.2 11B across zero-shot, one-shot, and fine-tuned settings, on two classification tasks spanning seven real-world disaster events. Our results show that zero-shot multimodal LLMs demonstrate reasonable generalization, while one-shot and five-shot prompting do not consistently improve performance. In contrast, fine-tuned open-source models—especially LLaMA 3.2 11B, substantially outperform all zero- and few-shot settings, particularly on complex tasks. We also observe that the LLaMA 3.2 11B model struggles with few-shot multimodal inputs involving multiple images, resulting in a performance drop in one-shot experiments compared to zero-shot. Fine-tuning LLaMA 3.2 11B on both single- and multimodal inputs achieves state-of-the-art results. Moreover, lightweight text-only models such as LLaMA 3.2 1B and 3B, when fine-tuned, can match or surpass previously best-performing approaches. These findings underscore the value of task-specific fine-tuning and offer a cost-effective path for applying optimized multimodal LLMs in real-time disaster response. Our code is available at this link .