Tavus Griffin: Real-Time Video AI Cuts Conversational Lag

The startup replaces multi-stage voice and text pipelines with a unified two-way video network; early trials fooled nearly half of callers in one-minute interactions.

On October 1, 2026, AI research startup Tavus unveiled Tavus Griffin, an interaction model engineered for direct, face-to-face video conversations. Instead of chaining separate services to listen, transcribe, process, synthesize speech, and animate lips, the system ingests and outputs synchronized video and audio in an end-to-end stream. For now, general availability is not open: the company released only an early research build, Griffin-Lite, to a restricted group of developers.

How the Tavus Griffin model replaces cascaded pipelines

Talking to an interactive digital avatar has long been clunky. In typical setups, user speech goes to an automatic speech recognition tool, text feeds into a large language model, the generated response passes to text-to-speech software, and an animation engine moves the mouth over a static image. That daisy-chained sequence piles up seconds of latency. It breaks the flow of conversation, makes natural interruptions nearly impossible, and strips out non-verbal reactions.

Griffin cuts out those middle layers entirely. According to technical documentation published by Tavus, the architecture takes raw camera and microphone feeds from the user and outputs matching video frames and sound directly. Running continuous conversational modeling with sub-second decision loops, the system monitors pauses, reacts to visual movements, and stops speaking when the human cuts in.

VideoFDB benchmark and the internal Turing trial

To evaluate the system, Tavus tested Griffin-Lite against NVIDIA’s Video Full-Duplex Benchmark (VideoFDB), an independent evaluation framework scored on a 1-to-5 scale. On the video generation track, Griffin-Lite recorded a score of 3.83, approaching the human baseline of 3.92 and outperforming a combined Gemini 2.5 and Anam setup at 2.80. On the perception track, the model reached 3.73 points against a human score of 4.20.

Tavus also ran a blind video Turing trial involving 54 participants in one-minute video calls. In that experiment, 48% of participants (26 callers) reported believing they were speaking with a real person. Under Phoenix-4.5, the company’s prior generation system, only 2.4% of 41 participants came away convinced.

Those numbers highlight notable progress, but they come with clear boundaries. The Turing test was conducted internally by Tavus without independent peer review, relied on a small pool of 54 people, and evaluated interactions capped at just 60 seconds. Whether that conversational realism holds during lengthy technical discussions or across unstable network conditions is still unproven.

Full-scene rendering and voice cloning

Rather than pasting mouth movements onto a fixed headshot, Griffin synthesizes every pixel in the video frame from a single reference image. The model computes room lighting, background shadows, and movement across the neck, torso, and hands. On the audio side, a diffusion engine can replicate a target voice using a 10-second reference recording.

Practical limits, compute costs, and access

Despite the technical demos, Griffin is not widely available. Tavus placed Griffin-Lite behind an approval-only waitlist for developers and has not announced a public release date for the full model on its commercial API.

The design choices show where conversational interfaces are heading: real-time interactive video without the stutter of legacy stacks. Yet the practical bottleneck remains hardware expense. Running diffusion networks to generate video frames live demands heavy GPU resources, and Tavus has not disclosed the compute infrastructure needed to run the architecture in production. The full model also remains restricted while the company implements safety controls against unauthorized voice cloning and deepfakes.