Tavus's Falcon cuts conversation video latency to 430ms

Tavus has published Falcon, a video diffusion model built for face-to-face conversation rather than scripted speech, and the number it leads with is latency. Tavus measured a mean audio-to-video delay of 430 milliseconds on a B200 GPU, which it says is about 50% below the fastest baseline it evaluated.

The pitch is that most avatar models animate a script, while Falcon is built to hold a conversation. It generates autoregressively, one latent at a time, which the team says gives finer control and lets the avatar interrupt, backchannel and react quickly. Falcon is driven by streaming audio and streaming text controls, so an upstream text-to-speech model and a vision-language model can change the avatar's nonverbal behavior mid-sentence.

What it does

Falcon generates 720p video in real time on a single GPU, including complex hand gestures, according to the technical write-up. Tavus says it is the first video-diffusion-based avatar model to follow both streaming text prompt changes and streaming audio, and reports state-of-the-art scores on FID, DOVER and THEval on a conversational test set despite the lower latency.

The work is the rendering layer behind Tavus's Griffin. On October 1, 2026, Tavus previewed Griffin-Lite, a full-duplex video-to-video conversation model, to invited testers, with the full Griffin model announced for later. Falcon is the piece that makes the delay low enough to feel like a conversation rather than a wait.

The limits

The authors are explicit about what is not solved. Falcon's temporal compression means one latent stands for eight frames, or 320 milliseconds, which sets the granularity of each generated chunk; reducing that unit while keeping real-time throughput is named as future work. Training with varying captions, the team notes, can still lead to occasional missed or delayed actions.

For creators, the practical read is that conversational video is moving from a pre-rendered asset to a live one. A 430-millisecond audio-to-video delay on a single high-end GPU is not consumer hardware, and Tavus has not published pricing or general availability for Falcon. But it moves the category closer to avatars you can talk with, not just talk at.

Falcon is described in Tavus's research write-up; Griffin-Lite reached invited testers on October 1, 2026, and Tavus has not announced general access.