Large Language Models (LLMs) excel at writing poems and summarizing research, but how do they handle the complexities of your medical chart? New research reveals that while LLMs hold immense promise for healthcare, they currently face significant hurdles in accurately interpreting clinical text, particularly when it comes to pinpointing specific medical entities within patient records. This critical task, known as token-level Clinical Named Entity Recognition (CNER), is essential for extracting precise details about diseases, symptoms, and treatments. The study investigated several leading LLMs, including both open-source models like LLaMA-2 and Meditron, as well as proprietary giants like ChatGPT-3.5 and ChatGPT-4. The surprising result? Even with advanced techniques like few-shot learning and access to vast medical knowledge bases, these LLMs struggled to match the precision of existing, more specialized clinical NLP systems. One bright spot emerged with Llama2-MedTuned, a medically adapted LLaMA-2 model. When fine-tuned on a rare disease dataset, it showed remarkable improvement, outperforming ChatGPT-4 and rivaling specialized clinical models. This finding indicates the potential of carefully trained, open-source LLMs to transform healthcare. Error analysis revealed the specific pain points for LLMs: difficulty in identifying precise entity boundaries and correctly classifying subjective symptoms. While this research highlights the current limitations of LLMs in clinical settings, it also points toward exciting future developments. As researchers explore new architectures, training methods, and prompt engineering techniques, the ability of LLMs to accurately decipher medical records could revolutionize diagnostics, treatment planning, and patient care.