Hippocratic AI announced Polaris 6 Ultra, a new generation of its healthcare language model that delivers six upgrades to how an AI care agent listens, thinks, and speaks. Polaris 6 Ultra is the first healthcare model with what the company calls bedside intuition: the ability to read a patient's emotional state, think mid-speech and decide what to say next in the same breath, without a pause the patient can hear.

"In healthcare, you don't get credit for reminding a patient. You get credit for the action they take," said Munjal Shah, Co-Founder and CEO of Hippocratic AI. "Polaris 6.0 was built from more than 323 million patient conversations, making it the only truly experienced frontier model for healthcare able to operate safely and autonomously to drive patient outcomes. That experience shows. Polaris 6.0 handles the healthcare edge cases of real life, it now has more human-like common sense and pacing, and it delivers the kind of empathy that persuades patients."

General-purpose models aren't built to address healthcare edge cases, or handle the realities of patient care. Traditional LLMs expect clean audio, reason before they speak, know only what is written down, read every line in one tone, and get slower as they get smarter. That doesn't work for healthcare.

"Everyone building voice AI for healthcare trades off model size for latency: speech-to-speech models are fast and shallow, reasoning models are capable and slow," said Subhabrata Mukherjee, Co-Founder and Chief Science Officer. "We refused that choice. We took the largest model that could be made safe and rebuilt the entire stack for voice, from the GPU numerics to the breath in the synthesized voice. Then, with full-parameter reinforcement learning at frontier scale, we gave Polaris 6 not just safety but common sense, intuition, and the warmth and realism of an experienced clinician."

With increased EHR integration and new flexible pricing, Polaris can now meet the needs of any healthcare client. Polaris 6.0 ships as Ultra (full constellation, all 31 supervisors, for clinical patient-facing conversations), Pro (sub-clinical workloads), Flash (high volume, subclinical), and Nano [administrative use cases], so customers can match coverage and speed to their specific needs.

Polaris 6 Ultra closes all six gaps at once:

• AI Intuition (v1) — more persuasive. The model generates an unspoken planning span before every sentence while the previous sentence's audio is still playing, reading the patient's mood and resistance and deciding what to bring up now and what to save. Time to first sentence stays at 278 milliseconds for the LLM. Every plan is logged against the sentence it produced and shown in the Safety Portal beside the transcript, so a clinician can audit, sentence by sentence, which risk the model noticed and which it deferred.
  
  
• Care Voice (v3) — more empathetic. The third generation of the Polaris voice is the first directed by the model rather than narrated. Breath, sighs, laughter, pauses and tempo are emitted as tokens from the plan, so pacing adapts to the patient's state and tone shifts before the words do. An in-house text-to-speech model pronounces 90% of 629 evaluated drug names correctly, versus 75% for the leading commercial TTS.
  
  
• Common Sense 2.0 — more common sense. Full-parameter reinforcement learning on the 671B primary model, end to end for the first time in a Polaris release, moved Hippocratic AI's internal healthcare-intelligence benchmark 18 points over its own base and 11 points past the best frontier model, with the largest gains in escalation and transfer, the first thirty seconds of a call, language switching, and call endings.
  
  
• Clinical ASR (CASR v3) — more inclusive. An in-house audio-LLM conditioned on the conversation and the chart hears patients who mumble, are recovering from a stroke, or are calling from a landline with the television on. On 2,778 clinician-reviewed production calls, misheard medications or symptoms fell from 1.19 to 0.20 per 100 calls, misheard numbers, names and emails fell from 0.57 to 0.00, and words invented on noise-only audio fell from 24% to 0.6%. Television speech attributed to the patient fell from 64.8% to under 3%.
  
  
• 7.4 trillion parameters — smarter makes it safer. The 671B primary model now runs inside a constellation of 30 specialist supervisors and 142 production monitors, up from 5.0 trillion parameters in Polaris 5, checking every turn in parallel across escalation and crisis response, medications, identity and consent, scheduling, privacy, and conversational quality.
  
  
• 1.3× faster — unlocks real-time thinking. Blackwell-native 4-bit numerics, routing-aware mixture-of-experts scheduling, first-token speculative decoding, and a serving engine Hippocratic AI owns deliver a 1.3× end-to-end speedup over Polaris 5 as patients experienced it, earned while the constellation grew and the model began reasoning mid-sentence. On byte-identical prompts on a 671B parameter LLM, the stack reaches the first sentence in 0.28 seconds versus 0.71 to 0.94 seconds for the four fastest commercial providers.

Across generations, the clinician-measured safety of the Polaris constellation has held at its ceiling since Polaris 4; with the latest constellation, the minor-harm rate in clinician output testing retains 0.00%. What has moved is the part patients feel: how natural, how perceptive, and how fast the agent is. Polaris 6 Ultra ships under the same criterion as every Polaris release: clinically tested, validated by licensed clinicians, and governed under human oversight before it reaches a patient.