Tavus has introduced Griffin, an artificial intelligence model capable of conducting real time conversational video calls with human level responsiveness. In a recorded video Turing test, the software convinced 48% of participants they were speaking to an actual human. Access remains strictly gated, with the Griffin Lite version restricted to private testing partners.
Traditional conversational systems process speech recognition, language models, voice synthesis, and visual rendering in separate stages. Griffin combines every layer into 1 single full duplex video to video model. The architecture evaluates the video stream multiple times every second. This allows the avatar to nod, offer short vocal interjections while the user speaks, stop talking when interrupted, and react immediately to visual movements on camera.
Rendering 720p video in 320 ms blocks, the system achieves an audio to visual latency of 0.43 seconds when running on NVIDIA H100 hardware. The platform also clones voices from a brief 10 second audio recording, allowing for rapid digital replica generation.
The breakthrough was validated through a video Turing test designed to measure conversational realism over live video calls. After a 1 minute conversation, 26 out of 54 human participants believed the AI was a real person. An earlier setup from Tavus convinced only 1 out of 41 participants under similar conditions. In third party benchmarks, Griffin secured 1st place among 15 competing systems evaluated in the NVIDIA VideoFDB test for video generation and scene comprehension.
Public access to the full model is currently blocked while safety evaluations continue. Tavus has released a smaller Griffin Lite variant solely to trusted enterprise partners. The team intends to monitor live interactions before considering a broader public rollout.
