OPEN LINK ↗
// 1h agoMODEL RELEASE
Tavus Griffin Nearly Passes Video Turing Test
Tavus introduced Griffin, a full-duplex video-to-video model that watches, listens, speaks, and renders expressive video continuously during live calls. In Tavus’s own one-minute study, 26 of 54 participants mistook Griffin-Lite for a human; the preview remains limited to select testers.
// ANALYSIS
The breakthrough is less about photorealistic avatars than about modeling conversation as a continuous audiovisual loop.
- –Griffin handles interruptions, backchannels, pauses, gaze, and turn-taking while it speaks.
- –Tavus reports 720p video generation in 320-millisecond chunks with 0.43-second audio-to-video latency on H100s.
- –NVIDIA’s VideoFDB leaderboard puts Griffin-Lite first in perception and near human-level generation quality.
- –The 48% result comes from a small, company-run study, so independent replication and longer conversations still matter.
- –Developers cannot broadly deploy Griffin yet; Tavus is withholding customer access while building disclosure and safety features.
// TAGS
tavus-griffinmultimodalvideo-genvoice-agentvisionspeechevaluation
DISCOVERED
1h ago
2026-10-04
PUBLISHED
1h ago
2026-10-04
RELEVANCE
9/ 10
AUTHOR
AI Revolution