Tavus Griffin — real-time video Turing-test avatar

Tavus's Griffin is a full-duplex "Human Interaction Model" that generates a conversational person — face, voice, body and background — frame by frame from a single reference image at 0.43s average latency.

Why it matters

What you could build with it

Does it hold up?

Too early to judge for builders: only the Griffin-Lite research preview exists for a small group of trusted testers, with no API, pricing or public release date; the 48% figure comes from Tavus's own 54-person study, though NVIDIA's independent VideoFDB scoring corroborates the generation quality.

Built with Tavus Griffin

• Frontier Now: Tavus launches Griffin, the full-duplex Human Interaction Modelfrontiernow.dev · Independent write-up: Griffin-Lite scored 3.83 on NVIDIA VideoFDB generation and 3.73 on perception (the highest on both leaderboards); only a trusted-tester research preview exists, with wider release gated on disclosure and safety work.
• Yomimono: Tavus Griffin AI fools 48% of video callers into thinking it is humanyomimono.id · Editorially reviewed summary of the GIGAZINE reporting: 720p video generated from one still image at sub-second latency, with the system split into a Continuous Conversational Modeling engine and an Audio-Visual Generation engine.
• StyleYourItems: Tavus unveils Griffin AI video model that fools 48% of usersstyleyouritems.com · Breakdown of the full-duplex architecture: Griffin reassesses conversation state at sub-second intervals, allowing it to nod, backchannel mid-sentence and stop the moment a user interrupts — versus relay-style avatar pipelines.

Learn more

technical deep dive →

First spotted on article: source.

More AI media tools