Tavus makes video agents meant to watch the caller and keep reacting while the caller talks. Its production system splits that work across separate models for turn-taking, perception and the face. Griffin, a research preview, is Tavus’s attempt to do it all in one model, and the published numbers suggest it behaves more like a person but answers more slowly.
1. Tavus and its video agents
Most of this series is about voice systems; Tavus works on faces. The San Francisco company was founded in 2020 and went through Y Combinator’s Summer 2021 batch. It sells a Conversational Video Interface (CVI), an API that gives an AI agent a real-time, talking human face and lets the agent look back at the user through the camera. Tavus calls these agents PALs. Hassaan Raza is co-founder and CEO, and Quinn Favret is COO. In November 2025 the company raised a $40 million Series B led by CRV, with Scale Venture Partners, Sequoia Capital, Y Combinator, HubSpot Ventures and Flex Capital taking part. It puts its total raised at about $64 million (Tavus company page).
Full-duplex is usually explained with two audio channels. Both sides can talk at once, so the system keeps listening while it speaks and decides, many times a second, whether to hold the floor or give it up. A camera and a screen on each side add two more channels. The user’s face, gaze and shared screen become input, and the agent’s face becomes output that stays on screen the whole time, including the stretches when the user is the one talking.
That last part is where many avatar systems stop being full-duplex. VideoFDB, a benchmark for audio-visual conversation from researchers at NVIDIA and David AI (May 2026), tested cascaded speech-to-avatar systems, where a voice model’s audio drives a rendered face. Its authors write that this design “fundamentally precludes the production of full-duplex nonverbal cues.” On the project page they add that a face driven by the agent’s own audio is effectively turn-based, with nothing to show during the user’s turn.
Tavus’s research since 2024 reads like a run of attempts at that problem. It built three model lines, Raven for perception, Sparrow for turn-taking and Phoenix for rendering, and on October 1, 2026 added Griffin, a research preview that tries to do all of it in one model. Between them they show what a full-duplex system needs once the agent can see and be seen.
2. The CVI pipeline
The CVI documentation lays out the production pipeline in six layers, three of them Tavus’s own. Raven reads the user’s expressions, gaze, background and shared screen. Sparrow handles turn detection and interruptions, deciding when the agent speaks and when it listens. Phoenix is the real-time face. In between sit speech recognition, a large language model (Tavus hosts low-latency models, or developers bring their own) and text-to-speech, where Cartesia is the default engine and ElevenLabs and Azure are options. Media runs over WebRTC, through Daily.
Tinted rows are Tavus’s own models. Media travels over WebRTC through Daily.
Architecturally this is a cascade. The words go from speech to transcript to language model and back to speech, and Tavus’s models work around that text path. The advantage of this design is that each part can be chosen and measured on its own, as the choice of language model and TTS engine already shows. The cost, as Tavus describes the layers, is that what the camera sees never reaches the decision about when to speak.
In November 2026 the troubleshooting guide lists sparrow-2 as the default turn-detection model and sparrow-1 as deprecated, and the perception docs keep raven-1 as the default with raven-2 available. Phoenix-4.5, released September 10, 2026, is live for every PAL. The CVI overview promises conversations in 42+ languages, a figure it ties to the supported TTS engines. It calls CVI fast, but it gives no number for the time between the end of the user’s speech and the start of the agent’s reply. For a product sold on real-time conversation, that is the most conspicuous gap in the docs.
3. Sparrow, the turn-taking model
Deciding who has the floor is a job a video agent shares with any voice agent. Tavus’s model for it has gone through three versions in under two years.
Sparrow-0 came out in March 2025 with Phoenix-3 and Raven-0. The launch post described a transformer-based turn-taking engine with sub-600 ms response times. Tavus’s own test in the later Sparrow-1 post put it at a 907 ms median with 13 interruptions, well over the launch figure. It is to Tavus’s credit that it published the worse number for its own older model.
With Sparrow-1 (January 13, 2026) the decision moved onto the audio itself. Tavus calls it audio-native and streaming. It reads continuous audio instead of transcripts, updates its state every 40 ms and predicts whether the system should listen, wait or speak. On Tavus’s own test of 28 samples, a small set, it answered a median 55 ms after the user stopped speaking and never interrupted. A silence-timeout baseline took a median of 1,002 ms and interrupted 59 times.
Sparrow-2 (August 27, 2026) is the current default. It still listens to raw audio. A Tavus encoder maps the signal into a shared space that carries semantic, prosodic, speaker and environmental information, and a six-layer causal transformer reads that stream every 10 ms. Tavus trained it on 16,000 real and 2,000 synthetic conversations. According to the post, it makes several calls at once that simpler systems split up: whether a turn has ended, whether an overlap is an interruption or a backchannel such as “mhm”, whether a voice in the room belongs to someone else, and whether a pause ends a thought or sits in the middle of one. The post also says it can wait through mid-thought pauses of up to 6 to 8 seconds and uses no silence timer.
| Model | Published | How it listens | Reported result |
|---|---|---|---|
| Sparrow-0 | March 2025 | Transformer-based turn-taking | Sub-600 ms response times (Tavus, 2025); 907 ms median in Tavus’s later Sparrow-1 test |
| Sparrow-1 | January 13, 2026 | Raw audio, a state update every 40 ms | Median 55 ms after the user stops, no interruptions (Tavus test, 28 samples) |
| Sparrow-2 | August 27, 2026 | Raw audio, six-layer causal transformer, every 10 ms | TurnBench test set: 0.930 recall, 0.084 false-positive rate, 615 ms |
Sparrow-2’s outside test came from TurnBench, a turn-taking benchmark released in August 2026. On its public development split (38 conversations, 7.3 hours of audio), Tavus reported preliminary figures of 92.4% end-of-turn recall and 97.4% interruption recall at a 632 ms median. The private test set, which TurnBench’s maintainers score, came back at 0.930 recall, a 0.084 false-positive rate and a 615 ms median. The leaderboard ranks systems by recall and does not qualify any with a false-positive rate above 0.15; on November 10, 2026 that put Sparrow-2 eighth. The seven systems above it are external submissions too. A score on a test set Tavus cannot see is worth more than its own dev-split figures. Eighth is solid, if not a win, and the research systems in the figure below, Voice Activity Projection among them, all sit further down.
End-of-turn recall, 0 to 1. The board ranks by recall; a false-positive rate above 0.15 does not qualify.
Voice Activity Projection, a research model, answers faster on the same test (368 ms) with fewer false positives, but it finds fewer turn ends (0.845 against 0.930). According to the TurnBench paper, listeners in smooth handovers start speaking a median 151 ms before the current turn ends, and the authors found no system that matches this without excessive false positives. A turn model that waits until it is sure arrives after that window, and one that guesses early cuts people off. With a 615 ms median and its long patience, Sparrow-2 sits on the waiting side. On a video call that is probably the better side to err on. A pause comes with a face that can keep showing attention, and an interruption is as rude on screen as it is on the phone.
4. Raven, the perception model
Raven-0, also from March 2025, gave the agent sight. It processed continuous camera input and tracked motion, gestures and eye contact.
Raven-1 (February 10, 2026) added audio. It reads tone, prosody, pacing and hesitation along with facial expression, gaze, posture and gesture, and instead of emotion labels or scores it writes plain-language descriptions. Tavus’s example is a speaker who sounds surprised and slightly skeptical, with a hint of irony. A description comes per sentence, so a long utterance gets several. Tavus gives three timing figures: audio descriptions in under 100 ms, combined audio-visual context that is never more than 300 ms stale, and a total pipeline under 600 ms.
Raven-2 is not on the research index; the docs describe it. They list emotional understanding from the user’s audio, better camera and screen understanding, and a visual history within the conversation, so the agent can refer back to something it saw earlier. Screen-awareness queries, and tools that fire on cues in a shared screen, require it. Raven-1 is still the default.
In CVI, what the camera sees reaches the conversation as text for the language model. That changes what the agent says next. It can notice that the user looks confused, or that a form on the shared screen has an empty field. Nothing in Tavus’s description says it changes when the agent speaks. Raven and Sparrow are separate layers, and the Sparrow-2 post describes the turn model’s input as audio, so a user who raises a hand or looks away to think is visible to the language model while the turn model hears only what the microphone picks up.
VideoFDB was built for cases like these. Its 237 clips, drawn from real video calls, cover 11 nonverbal conversational dynamics, and a language-model judge scores two tracks from 0 to 5: perception, which asks whether the agent read the situation, and generation, which asks whether the agent’s own audio and video hang together. On one item, a nonverbal interruption where the agent should yield within 1.5 seconds, the authors report Gemini 3.1 landing 1.7 seconds late.
5. Phoenix, the face
Phoenix is Tavus’s oldest line; the research index lists Phoenix posts from April 2024. Phoenix-3 (March 2025) was a “Gaussian-diffusion” renderer that animated the whole face. Phoenix-4 (February 18, 2026) combined diffusion with 3D Gaussian splatting, ran at 40 fps at 1080p by Tavus’s account, and added control over ten or more emotional states. While the user speaks it also generated what the post calls visual backchannels, such as nods, surprise and concern, and Tavus stressed that every listening frame is generated rather than looped from pre-recorded footage. The company says it trained the model on thousands of hours of human conversational data.
Phoenix-4.5 (September 10, 2026) changed what drives the face. Its animator, a streaming diffusion transformer, “listens to two audio streams, the PAL’s own speech and the user’s,” and turns both into motion. A renderer then paints the frame from one reference image. Phoenix-4 generated a face region of roughly 512 by 512 pixels; Phoenix-4.5 takes in the head, shoulders and torso in a single pass. Tavus reports 134 ms from audio to video without stating the hardware or the method (elsewhere the same post says under 130 ms), and 35 to 40 frames per second end to end on slower machines.
On a paired internal test set, Tavus reports lip-sync (AVSR) rising from 0.261 to 0.304, which it calls 16% better, and FID, a measure of how far generated frames sit from real ones, falling from 28.36 to 25.42. In head-to-head preference ratings, zero-shot Phoenix-4.5 scored an Elo of 1002 against 929 for a fine-tuned Phoenix-4. The post does not say how many raters took part, and without that an Elo gap is hard to weigh. All of these quality numbers are internal.
Even the recording for a custom face includes listening. The troubleshooting guide asks for 30 seconds of talking followed by 30 seconds of listening for a Phoenix-4 or Phoenix-4.5 replica, and the legacy Phoenix-3 asked for a minute of each.
Phoenix-4.5 is still the last stage of a cascade. Because its input includes the user’s audio, though, the face can respond while the user is still talking, which is what VideoFDB’s authors say cascaded avatars lack. Feeding the user’s audio into the animator sounds like a small change, but no other change in this story answers that criticism as directly. Whether those responses land at the right moments has no outside score yet, since the VideoFDB leaderboard has no entry for the CVI stack.
6. Griffin, the single-model preview
On October 1, 2026, Tavus announced Griffin, which it calls the first Human Interaction Model, a class of model for understanding and generating face-to-face, real-time interaction. The authors are Hassaan Raza, head of research Ioannis Patras and the Tavus Research Team. Only Griffin-Lite is available, as a research preview for select trusted testers. Griffin is not on the Tavus platform yet, and Tavus says a more powerful model will follow.
Griffin has two parts. A continuous conversational modelling engine takes in the user’s audio and video and makes its decisions at regular sub-second intervals, which Tavus calls mini-turns, rather than once per turn. It emits text plus control signals for speech and nonverbal behaviour, such as gesturing, looking away or changing emotion. An audio-visual generation engine follows those signals with two streaming generators. The speech generator, an autoregressive diffusion transformer, can clone a voice from about 10 seconds of audio. It runs on Tavus’s own codec, Tavec v1.7, which maps 48 kHz audio to 40 continuous values per frame at 100 frames per second, with packets as small as 10 ms. The video generator produces 720p video in real time, one latent at a time, from a single reference image.
In CVI, a turn model hears the audio, a perception model describes the video, a language model writes text and the face follows the sound. Griffin, as Tavus describes it, has one engine that sees and hears the user and decides at sub-second intervals both whether to talk and what the face and body should be doing. In principle, then, what the agent sees can move the floor.
Tavus reports an average video latency of 0.43 s on H100 GPUs (best 0.27 s, worst 0.59 s), about half that of the fastest method it compared against, AvatarForcing at 0.86 s. That covers the video generator only. VideoFDB measured how long Griffin takes to answer in a conversation.
On the VideoFDB leaderboard, Griffin-Lite scores 3.83 out of 5 on the generation track. The human ground truth scores 3.92, and the next entry, Gemini 2.5 driving an Anam Cara-3 avatar, 2.80. On the perception track Griffin-Lite scores 3.73 against a human reference of 4.20, ahead of the open MiniCPM-o 4.5 at 3.40 (3.44 in its audio-only baseline, with the video withheld). Griffin-Lite is slow to answer, though. Its median response latency is 1,892 ms on generation and 2,232 ms on perception, where the human references answer in 900 ms and 1,400 ms and MiniCPM-o 4.5 in 720 ms. The generation score is the striking number, with a single model nearly matching the human ground truth while the next system is far behind. The latency is the number to watch, because a reply that comes a second late is something people notice in conversation.
Overall score, 0 to 5, from the benchmark’s language-model judge.
Tavus’s headline number comes from a small study of its own. Participants held one-minute video calls about what they were looking forward to this year and were told only at the end that their partner was an AI. Of 54 people who spoke with Griffin-Lite, 26 (48%) said it was a real person. Of 41 who spoke with the CVI stack on Phoenix-4.5, 2.4%, or one person, said so, an unusually candid figure for a company to publish about its own product. Tavus calls Griffin-Lite the first model to pass a video Turing test “to the best of our knowledge.” The sample is small, the measure is what participants said they believed, and every call was on the same topic.
Tavus’s page raises the risk itself. It says the same properties that make these models natural could let them deceive a person into believing they are not talking to an AI, that more alignment and safety work is needed before release, and that it is building disclosure features. It gives no release date. For a model whose headline result is being mistaken for a person, the disclosure work matters at least as much as the next benchmark score.
7. What the camera changes, and open questions
Put next to each other, the two systems show what a camera adds to full-duplex conversation. A voice agent mostly has to decide when to take the floor. A video agent also has to read signals that make no sound, such as a frown, a glance at the screen or a raised hand. Its face is always on screen, so it needs to be doing something sensible during the user’s turn. And what the agent sees may need to change when it speaks, as well as what it says.
CVI splits these jobs across separate models. Raven turns the camera into text for the language model, Phoenix-4.5 is built to animate from both audio streams so the face can react while the user talks, and Sparrow-2 keeps turn-taking on audio. Griffin, as Tavus describes it, puts all of this inside one decision loop. On VideoFDB that loop scores closer to the human reference than any other entry and answers 0.8 to 1 second later than the human references. For someone building on Tavus today, that is the trade-off to weigh. CVI is available, and its parts can be checked one at a time. Griffin comes closer to a person, on VideoFDB and in Tavus’s own study, but it is a preview and slower to answer.
The outside evidence is uneven. For turn-taking, Sparrow-2 has a TurnBench test score from the benchmark’s maintainers. For the face, the production stack has only Tavus’s internal lip-sync and Elo figures, while Griffin-Lite has its VideoFDB rows. Tavus’s headline result is a 54-person study whose measure, being taken for a human, is the same property Tavus names as the safety risk. What is still unproven is whether Griffin’s lead holds outside a one-minute call on a single topic.
Four questions are still open:
- Can a single model that sees and hears close a gap of 0.8 to 1 second to human response times without losing what put it ahead on VideoFDB?
- Will Griffin get a paper? For now it is described on a Tavus page, and Sparrow-2’s TurnBench score is the only outside number for a model in the production stack.
- How will a model that nearly half the study’s participants took for a person tell people that it is not one?
- Will visual cues reach the production turn model, or only Griffin?
Sources
- Tavus: company information page (founding, team, Series B, total raised) and research index (model dates).
- Tavus docs: CVI overview, Perception, Troubleshooting.
- Tavus posts: Phoenix-3, Raven-0 and Sparrow-0 (March 6, 2025), Sparrow-1 (January 13, 2026), Raven-1 (February 10, 2026), Phoenix-4 (February 18, 2026), Sparrow-2 (August 27, 2026), Phoenix-4.5 (September 10, 2026), Griffin (October 1, 2026).
- TurnBench: paper (arXiv 2608.25218) and leaderboard.
- VideoFDB: paper (arXiv 2605.30256), project page and leaderboard.
Next: Thinking Machines Lab and its full-duplex interaction model