Polaris is a safety-focused large language model (LLM) constellation designed for real-time patient-AI healthcare conversations. Unlike previous LLMs in healthcare that focus on tasks like question answering, Polaris specifically targets long multi-turn voice conversations. The system comprises a one-trillion parameter constellation of several multi-billion parameter LLMs, including a stateful primary agent and specialist support agents. The primary agent drives engaging, patient-friendly conversations, while the specialist agents perform healthcare tasks such as medication adherence, lab results interpretation, and dietary guidance. The training protocol involves iterative co-training of the agents to optimize for diverse objectives, using proprietary data, clinical care plans, regulatory documents, and medical manuals. The models are trained to speak like medical professionals through organic and simulated healthcare conversations, enhancing capabilities such as rapport building, trust, empathy, and advanced medical reasoning. The evaluation of Polaris includes a comprehensive clinician evaluation with over 1100 U.S. licensed nurses and 130 U.S. licensed physicians posing as patients. The system demonstrates performance on par with human nurses in terms of medical safety, clinical readiness, patient education, conversational quality, and bedside manner. Additionally, the specialist support agents outperform a larger general-purpose LLM (GPT-4) and a medium-sized LLM (LLaMA-2 70B) in specific healthcare tasks. The architecture of Polaris is designed to handle real-time, voice-based interactions, addressing challenges such as voice quality, pitch, tone, response length, and interruptions. The system is modular, allowing for easy maintenance and continuous improvements. The constellation paradigm reduces latency and enables efficient multitasking among the agents, while maintaining robustness and safety through redundancy and specialized support agents.Polaris is a safety-focused large language model (LLM) constellation designed for real-time patient-AI healthcare conversations. Unlike previous LLMs in healthcare that focus on tasks like question answering, Polaris specifically targets long multi-turn voice conversations. The system comprises a one-trillion parameter constellation of several multi-billion parameter LLMs, including a stateful primary agent and specialist support agents. The primary agent drives engaging, patient-friendly conversations, while the specialist agents perform healthcare tasks such as medication adherence, lab results interpretation, and dietary guidance. The training protocol involves iterative co-training of the agents to optimize for diverse objectives, using proprietary data, clinical care plans, regulatory documents, and medical manuals. The models are trained to speak like medical professionals through organic and simulated healthcare conversations, enhancing capabilities such as rapport building, trust, empathy, and advanced medical reasoning. The evaluation of Polaris includes a comprehensive clinician evaluation with over 1100 U.S. licensed nurses and 130 U.S. licensed physicians posing as patients. The system demonstrates performance on par with human nurses in terms of medical safety, clinical readiness, patient education, conversational quality, and bedside manner. Additionally, the specialist support agents outperform a larger general-purpose LLM (GPT-4) and a medium-sized LLM (LLaMA-2 70B) in specific healthcare tasks. The architecture of Polaris is designed to handle real-time, voice-based interactions, addressing challenges such as voice quality, pitch, tone, response length, and interruptions. The system is modular, allowing for easy maintenance and continuous improvements. The constellation paradigm reduces latency and enables efficient multitasking among the agents, while maintaining robustness and safety through redundancy and specialized support agents.