Tavus Griffin: the first Human Interaction Model

What it claims

Griffin-Lite is a research preview of Tavus's Griffin, and it is the model that was actually evaluated. It is not available to customers. Only selected trusted testers can use it, through a request form. Tavus says it is withholding release until disclosure and safety features are ready. The page gives no weights, API, model sizes, compute figures or pricing. If you need to build something today, Tavus points you to its existing cascaded stack: Phoenix for the face, Raven for perception and Sparrow for turn-taking.

How it works

The system has two engines running at the same time.

• Continuous Conversational Modeling reads the user's audio and video. It decides at regular sub-second intervals, not once per turn, what to say and how to deliver it: tone, stance, expression and gesture. Because turn-taking depends on content rather than on the audio stopping, a pause to think is not treated as the end of a turn. The video input also covers the user's gaze, their surroundings and a shared screen.
• Speech generation is an autoregressive diffusion transformer. It clones a voice from about 10 seconds of reference audio, which it takes as a prefix, and emits one latent chunk at a time as controls arrive. It runs on Tavec, a convolutional autoencoder:
• Tavec encodes 48 kHz audio into 40 continuous values per frame at 100 frames/s, with no codebooks.
• Its decoder is causal and stateful, and outputs packets as small as 10 ms.
• One minute of speech becomes 6,000 latent frames, compared with 2.88M raw samples.
• Because the latents are continuous, training can use regression-based flow matching with no quantizer gradient tricks.
• Video generation replaces the 3D priors used in earlier Tavus models, which could not produce hand or body gestures or dynamic backgrounds. It renders the full scene, including chair, shadows and background, from one reference image. It was distilled from a large bidirectional many-step diffusion teacher in three stages: 1. Distribution Matching Distillation to a few-step, non-autoregressive student. 2. Teacher forcing to make that student autoregressive, generating one latent at a time. 3. Self-Forcing for long rollouts without drift, plus a mechanism on the history frames to improve stability.

flowchart LR
  U[User audio + video] --> C[Conversational engine]
  C -->|content + delivery| S[Speech generator]
  R1[~10s voice clip] --> S
  S -->|streaming audio| V[Video generator]
  C -->|nonverbal controls| V
  R2[Reference image] --> V
  V --> O[720p video out]
  S --> O

Evidence

• Video generator alone, compared with four unnamed published streaming diffusion models in an audio-to-video setup:
• Latency from audio to on-screen effect averaged 0.43 s on H100s, about half that of the next-fastest model.
• It ranked first on DOVER, FID and THEval.
• It ranked second on lip sync with LSE-C 7.27. Tavus argues that LSE-C rewards exaggerated mouth movement.
• NVIDIA VideoFDB, scored independently by NVIDIA across 15 models. Griffin-Lite is the only system evaluated on both tracks.

Takeover-rate alignment, which measures how well the timing of when the model speaks matches the reference, was 62.8% on generation and 73.8% on perception. Both are the highest on the leaderboard.

• Live study:
• Setup: participants came from an independent research platform. They were told they would have a one-minute call with another participant about what they were looking forward to this year.
• Main result: 26 of 54 (48%) believed their partner was human, compared with 2.4% (n=41) for Phoenix-4.5.
• Confidence: people who answered "human" averaged 79% confidence, and people who answered "AI" averaged 81%.
• Timing of suspicion: those who suspected an AI usually did so within the first 20 seconds.
• Ratings on a 7-point scale: natural 5.4, trustworthy 5.6, would talk again 5.8 (5.4 among those who identified it as AI), "really listening" 5.5. Conversational flow scored lowest at 4.9.

Limits

• Small study scope. The study used small samples, one-minute calls, a single topic, one persona, and participants who were not told an AI might be involved. That does not show the model holds up in longer or task-oriented conversations.
• Inconsistent prior-system number. The intro says previous Tavus systems peaked at a 2% pass rate, while the study section reports 2.4%.
• Unnamed baselines and missing details. The video baselines are not named. The page gives no model sizes, inference hardware requirements beyond "H100s", or cost.
• Untested model. Every result is for Griffin-Lite. The "more powerful" full Griffin has not been evaluated publicly.
• Stated risk. Tavus explicitly says the model can deceive people into believing it is human, and gives that as the reason for gating access.