- On Oct. 1, 2026, Tavus unveiled Griffin, a Human Interaction Model that can process video and audio while responding in real time; the company calls it the first model to pass a video Turing test.
- In a test involving 54 participants, 26 people, or 48%, said after a one-minute video call that Griffin-Lite was a real person.
- Griffin-Lite ranked first among 15 models evaluated in independent NVIDIA VideoFDB testing of full-duplex audiovisual interaction.
- Tavus has limited access to Griffin while it develops AI disclosure features and assesses risks arising from the model’s ability to impersonate people.
Tavus introduced Griffin as a new class of Human Interaction Model, or HIM, that can see, hear, speak and respond to a conversation partner’s nonverbal cues simultaneously. Unlike cascaded systems that recognize speech, generate a response and then voice it, Griffin makes decisions at regular sub-second intervals.
The model operates as a full-duplex video-to-video system. It can respond to pauses, gaze, facial expressions and gestures; listen and speak at the same time; interrupt or continue a conversation without losing context; and determine when to respond based on the length of a pause.
Griffin can also recognize objects and its surroundings and incorporate them into its responses. Tavus said the model adjusts its voice, facial expressions and gestures to fit the situation.
Real-time demonstrations and performance
In demonstrations, Griffin played Simon Says while tracking participants’ words and movements. It incorporated a peacock shown to the camera into a conversation’s storyline, helped solve a Rubik’s Cube by following hand movements and guided soldering work while keeping track of time.
The system has two core components. Continuous Conversational Modeling analyzes audio and video to decide when and how Griffin should respond, while Audio-Visual Generation converts those signals into speech and video.
According to Tavus, Griffin-Lite generates 720p video in real time using 320-millisecond chunks. On H100 GPUs, it recorded average latency of 0.43 seconds, which the company said was roughly twice as fast as the next-fastest method tested.
Griffin-Lite leads NVIDIA VideoFDB testing
Griffin-Lite ranked first among 15 models evaluated in independent NVIDIA VideoFDB testing designed to measure full-duplex audiovisual interaction, according to Tavus.
In the generation track, Griffin-Lite scored 3.83 points, compared with 3.92 for the human benchmark. It scored 3.73 points in the perception track.
The model also posted the highest alignment scores for when it begins speaking, reaching 62.8% in generation and 73.8% in perception.
Nearly half of participants identified Griffin as human
In a test involving 54 participants, 26 people, or 48%, said after a one-minute video call that their conversation partner was a real person. Tavus’ previous system — Phoenix-4.5 combined with Sparrow-2 and Raven-1 — recorded a rate of 2.4%, with one of 41 participants identifying it as human.
More than half of the Griffin-Lite participants did not consider during the conversation that they might be speaking with an AI, according to Tavus. Participants who suspected that their conversation partner was artificial generally did so within the first 20 seconds.
Access limited over impersonation risks
Griffin is available only to a limited group of testers as a research preview. Tavus said it must address security issues before a public release because the model’s ability to impersonate a person creates additional risks.
“The same properties that make Human Interaction Models powerful interfaces for natural communications between human and machine allow them to deceive a human into believing it is not AI,” the developers noted.
Tavus said it is working on AI disclosure features and additional risk-assessment procedures. The company plans a broader release of a more powerful version of Griffin after addressing the security issues.
Source: Incrypted
