Models · 1 October 2026 · 1 min read
Tavus Griffin-Lite: a video person who talks back
Previewed on 1 October 2026: a generated person on a video call, made from one photo and ten seconds of voice, open to a few testers.
By The Crewnema team
Tavus announced Griffin-Lite on 1 October 2026 as a research preview of what it calls its first human interaction model. It is a full-duplex video model: it watches and listens to the person on the call and answers as a generated person, in real time. These are the points from Tavus’s announcement.
What Tavus announced
- A whole scene from one reference image: face, arms, fingers, the chair and the background.
- A cloned voice from about 10 seconds of audio.
- Expressions, gestures, nods and reactions to being interrupted.
- Video latency averaging 0.43 seconds on H100 GPUs.
- Available only to selected testers; Tavus says more safety work is needed before release, including ways to make clear the person is not real.
Our view
Tavus is right to hold this back, and says why itself: a generated person convincing enough to pass as real is a risk as much as a feature. For a business it points at a live spokesperson for a website or a shop kiosk, not at adverts. For adverts we would keep to the simpler rule: a real person’s face appears only with their permission, and a viewer should never be misled about whether a person is real.
Side by side
In Crewnema
Griffin-Lite is not available to use and is not on Crewnema’s roster. When a video uses a real person’s photo, Crewnema asks for that person’s permission before it is posted.
Sources
Make your first video
Crewnema is in private beta and lets businesses in a few at a time. Request access and we will write when yours is open.