Large Language Models hold the potential to revolutionise the field of radiology by automating radiology tasks and enhancing clinical decision-making. However, the substantial computational resources required for training LLMs with a large number of parameters create barriers to their use both in radiology-related research and potential clinical applications. This study evaluated the performance of two smaller large language models (Meta-Llama-3-8B-Instruct and Mistral-7B-Instruct) in comparison to a newly introduced state-of-the-art multimodal model, LLaMA 4 Scout, for the task of impression generation in radiology. Experiments were conducted with and without in-context learning to ascertain its impact on model performance. Results show that in-context learning considerably improves the performance of all models, allowing them to achieve competitive results in the task of impression generation. These findings underscore the capabilities of small open-source LLMs and highlight the potential of advanced multimodal models for radiology-specific NLP tasks.