aivideo.zone Tavus Griffin-Lite: a video model that watches you and talks back, in 0.43 seconds
Griffin at TavusGriffin-Lite research preview and waitlist
Most AI video is a clip you ask for and wait for. Tavus, the company behind the Phoenix talking-head models, has shown something built the other way round: a model that is on the call with you. Griffin-Lite was announced on 1 October 2026 as a research preview.
What it is
Tavus calls it a Human Interaction Model: a full-duplex video-to-video system that takes your camera and microphone in and produces a face and a voice out, continuously, so that it can listen, interrupt, be interrupted and react while you are still speaking. Output is 720p in real time.
The latency
Response time averages 0.43 seconds on a single H100, in a range of 0.27 to 0.59 seconds, measured from the end of the user's turn to the start of the model's. Human conversation runs at a few hundred milliseconds, so this is in the range where a pause reads as thinking rather than lag.
The study
Tavus had 54 participants talk to it; 48 per cent judged it human, against 2.4 per cent for the company's own Phoenix-4.5, its current production model, in the same test.
Availability
A waitlist at griffin.tavus.io. It is not available to Tavus customers and the company gives no date; the preview is to show the direction.
Read the two numbers together. A 0.43-second turn and a 48 per cent human rating are the same fact from two sides: the thing that makes a generated face feel like a person is not the skin but the timing, and every talking-head product until now has had a second or more of dead air that no amount of rendering could cover. The honest caveats are that 720p on one H100 for one conversation is an expensive way to run a call centre, and that a test where half the participants were fooled is also a test where half were not. It is a direction, which is what Tavus says it is.
Tavus Griffin-Lite: a video model that watches you and talks back, in 0.43 seconds: questions
What is Tavus Griffin-Lite?
A research preview of a full-duplex video-to-video model that holds a face-to-face conversation in real time: 720p output, about 0.43 seconds of response on an H100. Announced 1 October 2026, waitlist only.
Can I use Griffin-Lite now?
No. It is a research preview with a waitlist at griffin.tavus.io; Tavus has given no release date and it is not part of the company's product.
Source: Tavus, Griffin checked against the source
More AI video news
LTX-2 v1.5.0
Lightricks released v1.5.0 of the LTX-2 code that runs the open LTX-2.5 AI video model on 18 October 2026: image conditioning now says whether a frame is replaced or only guides, keyframe interpolation moved into the main two-stage pipeline, and fp8 matrix maths speeds up GPUs of compute capability 8.9 and newer. Outputs are no longer bit-identical to earlier releases.
Runway MCP: sound and voice
Runway's MCP server, which lets an AI agent drive Runway's tools, gained sound in mid-October 2026: Mirelo SFX 2.0 writes a separate effects track for an AI video clip from a description, and Fish Audio Drama 3 turns a script into a voiceover in seven voices. BiRefNet v2 background removal arrived the same week. All three need a paid plan.
Helix World preview
Noiz AI and HKUST have released Helix World preview, an interactive audio-visual world model: 768×512 AI video at 24 frames a second generated live on a single H800 as you steer with WASD or the arrows, with 48 kHz spatial stereo sound. Built on LTX-2.3, weights on Hugging Face, demo at helixworld.org.