Tavus · 12× · tolest 03. Okt. 2026
Griffin is Tavus's first "Human Interaction Model" (HIM), introduced on October 1, 2026. It is a full-duplex video-to-video model that simultaneously sees, hears, speaks, and responds, adapting facial expressions, tone, and timing in real time to the conversation flow—rather than chaining separate modules for speech recognition, language model, speech synthesis, and avatar animation in sequence. In an internal Tavus study, 48% of test subjects believed the model was a real human after a one-minute video call; on NVIDIA's VideoFDB-Benchmark, Griffin-Lite achieved a top score of 3.83 out of 5 points (human reference value: 3.92). Currently, only a preview version called "Griffin-Lite" is available as a Research Preview for selected testers; there is no public
Features
| Real-Time Streaming | Full-duplex video-to-video streaming; generates 720p video in real time in 320 ms chunks (25 fps) |
| Latency | Audio-to-video latency averaging 0.43 seconds on H100 GPUs |
| License | Closed research preview program for selected trusted testers only (access via request form) |
| Platform | Not yet integrated into the Tavus platform; no public model ID, API, or endpoint available |
| Price | Not published; Griffin-Lite is only available as a free research preview to selected testers |
| Release Date | October 1, 2026 (announcement); only preview version "Griffin-Lite" available, no date for full release |
| Voice Cloning | Can clone a voice from approximately 10 seconds of reference audio |