Tavus Griffin: The Human Interaction Model, Explained

Tavus Griffin is a Human Interaction Model — HIM, the vendor's own category name — and the most useful thing to establish about it first is what it is not. It is not a video generator. A video model takes a prompt, a still frame or a clip and returns a finished file; the file is the product. Tavus Griffin takes a person sitting in front of a camera and holds a conversation with them, generating the face, the voice and the nonverbal behaviour in the same moment it is listening and deciding, continuously, for as long as the call runs. The only version of it anyone outside the company has touched is Griffin-Lite, a research preview whose page carries a date of 1 October 2026, open to select trusted testers through a request form, with no published price, no model identifier in Tavus's API documentation and no endpoint. This page is the standing answer to what the model is, what it runs on, what has actually been measured, and who can use it.

The class question, because it has been getting answered wrong

Search for Tavus Griffin and a reader is as likely to land on a comparison against a text-to-video or image-to-video model as on anything about conversation. That framing is worth correcting rather than working around, because everything downstream of it — what the model is good for, what its scores mean, whether its latency is fast or slow — depends on which class you file it under. The two classes are not variations on a theme. They accept different things, produce different things, and are measured on different instruments.

• Video-generation models — input is a prompt, an image or a clip; output is a file of a fixed length; the interaction ends when the render does. Benchmarks score the pixels.

• Tavus Griffin — input is a live person, their audio and their video, continuously; output is a stream of speech and video that never has to end; the interaction is the product. Benchmarks have to score the exchange.

Griffin is also, in the vendor's framing, a rejection of the usual way an interactive agent is assembled. Tavus explicitly contrasts it with cascade systems that chain speech recognition, a language model and speech synthesis one after another, waiting for the user to finish speaking before beginning to respond. Griffin is described as full-duplex video-to-video: perception, the decision about when and how to respond, and the generation of the response all run at the same time, and they keep running while the model is speaking.

There is a signal in the wording readers actually type. One of the recorded queries for this model is "tavus griffin human interaction model", which is a searcher describing the class before they have found a page that names it. It is the right description. Tavus coined the term, abbreviates it HIM, and titles the model's own page The First Human Interaction Model. So the short answer to what Tavus Griffin is: a real-time conversational avatar system, classed by its maker as a Human Interaction Model, which is a different product category from a video model rather than a better one.

What it takes in and what it produces, live

The inputs are a live person's audio and video, plus a system prompt, the conversation history, a reference image and a reference audio clip, and the ability to call a tool mid-conversation. On the output side the conversational model does not emit only text: it emits text plus a set of control signals that direct the speech and video generators, which is how tone, stance, facial expression and gesture stay tied to the decision that produced them.

On the generation side, Tavus says the architecture produces 720p video in real time, one latent at a time, accepting streaming controls so the conversation model can direct nonverbal behaviour such as gesturing, looking away or changing emotional register. The speech side emits audio packets as small as 10 ms, and a voice can be cloned from roughly ten seconds of reference audio. Every one of those figures is the vendor's own description of its own system, on its own page.

Latency is the one number in that group the vendor publishes as a measurement rather than a design goal. Against four published streaming diffusion models, in an audio-to-video setting where each system is given speech and a reference image and must produce the talking face, Tavus reports Griffin-Lite averaging 0.43 seconds of true audio-to-video latency on H100s, with a best case of 0.27 s and a worst case of 0.59 s. For scale, the vendor's own figures for the next-fastest method in that comparison average 0.86 s, and the slowest averages 1.45 s. Tavus ran that comparison against baselines Tavus selected, so treat it as a vendor-reported figure with a specific protocol attached, not as a neutral ranking.

One thing the vendor does not publish is the video generator's frame geometry. There is no frames-per-second figure on the page and no statement of how many video frames one latent covers, which means any specific chunk duration you have seen quoted for Griffin — 320 ms, or an eight-frame latent, or a stated compression ratio — is not on Tavus's page today and should not be repeated as though it were. What is stated is 720p, one latent at a time, and the 0.43-second audio-to-video average.

What the pipeline is actually made of

Tavus describes Griffin as a two-part system, and both parts are documented in enough detail to check.

The first is a Continuous Conversational Modeling engine. It perceives incoming audio and video, decides when and how to respond, and emits the text and control signals. The design point is the interval: it makes conversational decisions at regular sub-second intervals rather than once per turn, which the vendor calls mini-turns. At each one it assesses the state of the conversation and can take the floor, hold it, yield it, backchannel without interrupting, or stop when interrupted. Because the decision is continuous rather than turn-boundary-driven, a pause for thought is not treated as the end of a turn. Griffin also perceives video where most interactive systems perceive audio alone, which is what gives it gaze, facial expression, the user's environment, and anything the user holds up or shares to the camera.

The second is an Audio Visual Generation engine, itself two architectures — streaming speech generation and streaming video generation — converting those control signals into sound and pixels. The speech model is a fast autoregressive diffusion transformer that takes an encoded reference clip as a prefix and generates new speech progressively, one latent chunk at a time, as increments of output and control arrive, rather than waiting for the whole utterance to be composed before it starts speaking.

Underneath the speech side sits a codec the vendor calls Tavec v1.7, and this is where the published numbers get specific. It is a convolutional autoencoder mapping 48 kHz audio into a continuous latent of 40 values per frame at 100 frames per second, with no codebooks. Its decoder is fully causal, so it carries state across chunks and produces one seamless waveform from streamed latents. The compactness is the point: a minute of speech is 6,000 latent frames against 2.88 million samples of raw audio, which is what makes the sequence short enough for an autoregressive transformer to predict faster than real time. Tavus reports 2,193× real-time decoding, a ViSQOL MOS-LQO of 4.242, and 28.42 million decoder parameters, with the throughput measured on an H100. All vendor-reported.

The video side arrives at the same place from the other direction. Tavus's earlier models leaned on extensive 3D priors, which limited what they could express; the vendor's stated requirements for Griffin's video generator are fast response to each incoming audio chunk, tolerance for controls that change quickly, and stable identity, lip sync and visual quality over long generations. The route there was a three-stage distillation of a large bidirectional many-step diffusion teacher. Distribution Matching Distillation produced a few-step student that was fast but still generated a whole video in one chunk. Teacher forcing converted that student into an autoregressive model producing one latent at a time. Self-Forcing then trained it on its own generated history, with an added mechanism acting on the history frames, so that long rollouts do not drift — which is the failure mode this class of model normally has, a face that degrades or a background that wanders over a conversation that runs for minutes.

What has been measured, and by whom

Almost every number above is Tavus measuring Tavus. The exception, and the reason it carries weight, is NVIDIA's VideoFDB — a benchmark for full-duplex audio-visual conversation that NVIDIA built, and that NVIDIA states it conducts independently using the published metrics and its own judge. It scores a model on two tracks, generation and perception, each with its own rubric and its own leaderboard, on a zero-to-five scale.

• Generation — what the agent produced, audio and video together, in response to a dynamic event in the user's stream. Griffin-Lite scored 3.83 against a human ground-truth reference of 3.92; the next-highest system scored 2.80. Takeover-rate alignment, which measures how closely the model's decisions about when to speak match the timing in the reference conversations, was 62.8% for Griffin-Lite and 78% for the human reference, with a timing figure of 1,892 ms.

• Perception — audio-visual in, audio out, scoring whether the agent read the moment correctly. Griffin-Lite scored 3.73 against a human reference of 4.20, with takeover-rate alignment of 73.8% and 2,232 ms of timing. Tavus describes it as the most perceptive model on the benchmark among the 15 evaluated.

Two caveats belong on those before anyone repeats them. The nearest baseline on the perception track is listed as MiniCPM-o 4.5 at 3.44 — but that is the audio-only (A2A) row; the same system's full audio-visual row is 3.40. Griffin reads video, so the honest comparison is against the audio-visual rows, which puts the gap at 0.33 rather than 0.29, and part of what the gap measures is simply the value of having eyes. And a 0.09-point gap to a human reference on a five-point scale scored by a language model is better read as "close to human on this rubric" than as "at human level".

The other headline number is Tavus's own. In a study run by the company, participants recruited through an independent research platform were told they would be matched with another participant for a one-minute video call; their partner was a PAL running Griffin-Lite. Of 54 participants, 26 — 48% — believed their partner was a real person. On the same protocol, the company's previous stack built from Phoenix-4.5, Raven-1 and Sparrow-2 produced 2.4%, or one participant in 41. Both groups were confident: 79% among those who said real, 81% among those who said AI. More than half said the possibility had not crossed their mind at all, and those who did suspect mostly did so within the first twenty seconds. On a seven-point scale the model averaged 5.4 for seeming natural, 5.6 for trustworthy, 5.8 for wanting to talk again, 5.5 for feeling listened to, and 4.9 — the lowest of the five — for whether the conversation flowed naturally.

That study is a vendor-run study with a described protocol and a stated sample size, which is more than most model launches offer and less than an independent replication. No third party has reproduced the face-to-face protocol, and Tavus Griffin has no entry on an independent video-model leaderboard. The two VideoFDB track scores are the only figures on this page that an organisation other than Tavus produced.

Access and price: what exists today

This is the shortest section on the page because there is almost nothing in it. What is published, on the vendor's own material:

• Status — Griffin-Lite is a research preview, and Tavus states it will not be available for customer use at this time, though it is available for select trusted testers.

• Access — by request form, with a wider release of a more capable model described as following.

• Platform — the vendor's own closing line says Griffin is not on the Tavus platform yet, and will come once the company has worked out how to release it safely.

• Price — none published. Tavus's pricing page covers the PAL product and the Phoenix, Raven and Sparrow stack; it says nothing about Griffin, and there is no per-minute, per-second or per-request figure anywhere on the vendor's site.

• Endpoint and identifier — none. Griffin does not appear in Tavus's API documentation index, whose models page documents Phoenix for face rendering, Raven for perception and Sparrow for turn-taking. There is no model string to call.

The vendor states it anticipates releasing Griffin very soon after the safety work it describes is addressed. That is a stated intention, not a date, and it is the only forward-looking statement on the topic.

What Tavus Griffin is not for

The gap between the benchmark result and the release status is the practical content of this page, so it is worth being blunt about the things the model cannot currently be used for.

It is not something to build a customer-facing path on this quarter. A deployment needs a model identifier, a price, a rate limit, a region story and some form of availability commitment, and none of those exist. A team that plans around Griffin today is planning around a request form.

It is not a video generator, and it should not be dropped into a clip pipeline on the strength of the numbers. Its scores come from a full-duplex conversation harness with its own rubric and its own judge; video-model leaderboards score short clips by pairwise preference on prompts. The two instruments do not produce comparable numbers, which means Tavus Griffin has not been benchmarked head-to-head with any video model at matched effort and harness, and no ranking between them should be inferred from what is published.

It is not, by the vendor's own account, safe to release openly yet. Tavus writes that the same properties that make a Human Interaction Model a better interface also make it better at deceiving a person into believing it is not an AI, that further alignment and safety work is required, and that disclosure features and outside safety evaluations are in progress. Given that nearly half a live sample could not tell, and that the participants who could mostly worked it out inside twenty seconds, the disclosure question is not a footnote to the model — it is the gate on it.

What is buyable today is the rest of the stack around it. A conversational video pipeline is a chain — speech recognition, a language model, speech synthesis, and then the rendering layer that makes it a face — and the first three are commodity, interchangeable, and already routable. Those run behind one OpenAI-compatible endpoint at OrcaRouter, with over two hundred models on a single key, provider list price passed through at 0% markup so a vendor price change lands the same day rather than after a contract cycle, automatic failover that reroutes when a provider degrades — which matters more in a live call than in a batch job, because the failure is audible — plus a routing DSL for expressing which model should handle what, and model fusion for combining outputs. The generation layer is the part that is still a research preview; the layers above it are the part a team can assemble now. Tavus Griffin does not appear on OrcaRouter's live model list, which is consistent with a model that has no published endpoint and no identifier to route to.

What would change this page

Three things, and they are specific enough to watch for. A model card with a model string and a price would move Tavus Griffin from research result to procurement question. An independent replication of the face-to-face protocol — a second lab running the same one-minute call with its own participants — would do more for the 48% than any further vendor study can. And a long-session report from someone outside Tavus would test the one claim the architecture is built around and no benchmark covers: that Self-Forcing and the history-frame mechanism actually hold a face and a background together over a conversation that runs for minutes rather than seconds. Until one of those lands, the two VideoFDB tracks are the strongest evidence available, and they come with a 0.09-point gap to human on generation and 0.47 on perception.