Imagine an AI that not only holds vast medical knowledge but also understands the nuances of clinical practice. That's the promise of MedCare, a groundbreaking approach to building medical Large Language Models (LLMs). Current medical LLMs excel at either knowing a lot or following specific clinical guidelines, but rarely both. This is like having a brilliant medical student who aced the exams but struggles to apply that knowledge in a real hospital setting. MedCare tackles this challenge head-on by separating the learning process into two stages. First, it acts like a knowledge sponge, soaking up diverse information from various medical sources. This stage builds a strong foundation of medical understanding, like giving our student an intensive internship. Then, in the second stage, MedCare fine-tunes its knowledge to align with specific clinical tasks, like training our now-experienced student in a specialized field. This two-pronged strategy allows MedCare to master both medical knowledge and clinical practice, a feat that previously eluded many LLMs. Testing across 20 diverse medical tasks, including knowledge exams and clinical scenario evaluations, shows that MedCare outperforms other LLMs, especially with fewer parameters, which implies better efficiency and less risk of hallucinations. It’s like our star student graduating top of their class, ready to take on the complexities of real-world medicine. While MedCare holds immense potential, it's important to remember that it’s not intended for direct patient care just yet. It still faces the typical LLM challenges of potential biases and occasional inaccuracies. Think of it as a powerful tool for medical professionals, offering insights and assisting in decision-making. Further work aims to refine this two-stage learning process and further enhance the alignment between medical knowledge and practical application, paving the way for future iterations of clinically relevant medical LLMs.