Tavus

Tavus is releasing Phoenix-4.5, the newest version of its real-time human rendering model that powers the digital replicas, or "pals," inside its Conversational Video Interface, a stack developers use to build AI agents that speak, see, and hear in live face-to-face conversation . It is aimed at teams building tutors, healthcare intake, and companion products where an on-screen person has to hold attention through a full session, not just a demo clip. This launch matters because it is the first Phoenix model that can render a new identity zero-shot, showing a usable preview immediately while fine-tuning finishes in the background in roughly two hours instead of four. The technical shift is a rebuild of the renderer. Phoenix-4 relied on a 3D Gaussian splatting approach focused on the face , while 4.5 is a GAN-based network trained to generate more of the person at once, extending movement through the shoulders and torso and removing the seam artifacts that appeared when motion pushed past the facial region. It also handles the messy parts of real appearance that broke earlier models, including long hair falling over shoulders, glasses, earrings, and headbands, so people no longer have to change their look to get a clean replica. Tavus was founded in 2020 and is led by CEO Hassaan Raza and COO Quinn Favret , with backers including Sequoia, Scale, Y Combinator, and HubSpot . Phoenix-4.5 slots in alongside the company's Raven perception model and Sparrow turn-taking model as the visual layer of that stack, and it is the piece most directly responsible for whether a conversation with an AI agent feels like talking to a present person or watching a puppet.