Modern artificial intelligence has the potential to transform healthcare, yet integrating large language models into clinical practice presents challenges, especially concerning data privacy. This paper introduces Meditron 3, a suite of open-source medical language models developed in close collaboration with clinicians and based on Llama 3.1. Meditron 3 offers models with 8B and 70B parameters designed to support tasks like differential diagnosis and medical reasoning. Trained on a comprehensive medical corpus-including textbooks, anonymized patient records, and internationally recognized guidelines-these models aim to enhance medical services while respecting patient confidentiality. We conduct extensive evaluations using benchmarks such as MedMCQA and MedQA. Our 70B parameter model demonstrates promising results, outperforming Med-PaLM 2, and surpassing the base Llama 3 70B model by over 2% on average, while trailing GPT-4-Base by approximately 3%. By incorporating continuous feedback from medical professionals, Meditron 3 seeks to facilitate the safe and effective integration of AI into clinical settings. We publicly release Meditron 3 to encourage further development and adoption of artificial intelligence in healthcare. Ongoing work focuses on enhancing model reliability and ensuring actual compliance with medical standards to support widespread clinical use through the MOOVE.