Menlo Park – November 02, 2026 -- Hippocratic AI has launched Polaris 6 Ultra, a healthcare language model that runs a 7.4-trillion-parameter safety constellation 1.3 times faster end to end than its predecessor, Polaris 5, as patients experience it.
The company says the model is the first in healthcare to read a patient's emotional state and plan its next sentence while the previous one is still being spoken. Polaris 5 ran at 5.0 trillion parameters.
Polaris 6 Ultra reaches the first sentence in 0.28 seconds against 0.71 to 0.94 for rivals
On byte-identical prompts on a 671B-parameter LLM, Hippocratic AI's serving stack reaches the first sentence in 0.28 seconds. The four fastest commercial providers took 0.71 to 0.94 seconds. Time to first sentence for the LLM stays at 278 milliseconds.
The company credits Blackwell-native 4-bit numerics, routing-aware mixture-of-experts scheduling, first-token speculative decoding and a serving engine it owns. The speedup came while the constellation grew and the model began reasoning mid-sentence.
A 671B primary model now runs inside 30 supervisors and 142 production monitors
The constellation checks every turn in parallel across escalation and crisis response, medications, identity and consent, scheduling, privacy and conversational quality. In clinician output testing, the minor-harm rate remains at 0.00%, a ceiling the company says safety has held since Polaris 4.
Full-parameter reinforcement learning on the 671B primary model, applied end to end for the first time in a Polaris release, lifted Hippocratic AI's internal healthcare-intelligence benchmark 18 points over its own base model. It also beat the best frontier model by 11 points. The largest gains came in escalation and transfer, the first thirty seconds of a call, language switching and call endings.
Clinical speech recognition cuts misheard medications from 1.19 to 0.20 per 100 calls
The in-house audio model, CASR v3, is conditioned on the conversation and the patient chart. On 2,778 clinician-reviewed production calls, it produced these results:
- 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.
- 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%.
Care Voice v3 pronounces 90% of 629 drug names correctly versus 75% for leading commercial TTS
The third-generation voice is directed by the model rather than narrated. Breath, sighs, laughter, pauses and tempo are emitted as tokens from the model's plan, so pacing adapts to the patient's state.
Each unspoken plan is logged against the sentence it produced and shown in the Safety Portal beside the transcript. Clinicians can audit which risk the model noticed and which it deferred.
CEO Munjal Shah said Polaris was built from more than 323 million patient conversations and that in healthcare, "you get credit for the action they take."
Four model tiers and deeper EHR integration widen the addressable client base
Polaris 6.0 ships in four tiers:
- Ultra: the full constellation, for clinical patient-facing conversations.
- Pro: sub-clinical workloads.
- Flash: high-volume, subclinical use.
- Nano: administrative use cases.
The company is pairing the tiers with increased EHR integration and new flexible pricing. Hippocratic AI says its agents have completed more than 250 million patient voice interactions across more than 300 live clinical use cases, with zero safety issues in production. Polaris 6 Ultra ships clinically tested, validated by licensed clinicians and governed under human oversight before it reaches a patient.