Griffin by Tavus

Griffin listens while it speaks. It can hold the floor, yield it, or stop when you cut in, and it keeps tracking the conversation during its own response. Tavus built this as a continuous loop rather than a turn-by-turn handoff, which is why a pause in your sentence does not automatically end your turn.

The model generates every pixel in the frame from a single reference image. That covers the face, hands, and the chair the person sits in, along with shadows and the background. Most AI avatar systems animate a cropped head, so a moving hand or a shifting body breaks the illusion. Griffin does not have that seam.

Video comes out at 720p while the call is running, one latent at a time. That streaming output is the core of the real-time video AI claim. Tavus measured an average of 0.43 seconds to generate the smallest video chunk on an H100, with a best case of 0.27 seconds. The number that matters on a call is the gap between your interruption and the face reacting, and half a second is short enough to read as attention.

Griffin takes video input, not audio alone. It picks up gaze, facial expression, gestures, and objects you show the camera, and it can read a shared screen. That is how it reacts when you hold up a broken part or look away to think, instead of answering the transcript of your words.

The model changes tone, expression, and gesture with the context of the conversation. It laughs at the right moment, shifts when the subject turns serious, and uses non-verbal signals to acknowledge what you said. Tavus trained these behaviors as part of the same system that decides what to say, not as a layer added afterward.

Griffin's speech generator clones a voice from about ten seconds of reference audio. It runs on Tavec, Tavus's own audio codec, which streams in chunks as small as 10 milliseconds. For anyone building a branded AI avatar, that short sample requirement removes a recording session from the setup.