Tavus drops three AI models for real-time conversational video

The uncanny valley in AI video calls rarely comes from pixels. It comes from the awkward pause after you finish speaking, or an avatar staring blankly while its audio plays out of sync.

Conversational video platform Tavus wants to kill that awkwardness. On March 6th, the company rolled out three dedicated AI models designed to overhaul face-to-face AI interactions: Phoenix-3, Raven-0, and Sparrow-0, all made available immediately via API.

Instead of shoving every task into a single video generator, Tavus separated human conversation into three distinct mechanics: look, vision, and timing.

Breaking down the trio: Face, eyes, and conversational flow

First is Phoenix-3, a full-face animation model focused on micro-expressions and precise lip sync. It is built to maintain identity fidelity while keeping up with continuous, context-aware emotional shifts as the conversation unfolds.

Next is Raven-0, which acts as the avatar's visual perception engine. Raven-0 interprets ambient context and real-time visual cues, allowing the agent to read on-screen text, track hand gestures, and read human emotions on the fly.

Then there is Sparrow-0, and this might be the most critical piece. Sparrow-0 handles speech rhythms and conversational turn-taking, solving the classic problem of AI either interrupting you or leaving dead air.

In turn-accuracy benchmarks, Sparrow-0 posted a Mean Absolute Error of 0.3989, crushing the previous state-of-the-art mark of 1.7467.

Inside Tavus's conversational video operating system, these models link together into a unified chain-of-thought pipeline. To show off what that looks like, Tavus demoed Charlie, an AI agent capable of chatting fluidly, searching the web, analyzing shared screens, and generating images mid-call.

Why it matters

For builders creating AI sales reps, digital coaches, or customer support agents, this is a massive leap forward. Developers can now plug in via API to spin up interactive avatars that do not feel like pre-recorded interactive voice response bots.

We think Tavus's modular approach is brilliant. While much of the industry keeps obsessing over cinematic 4K video rendering, real-world human connection actually hinges on pacing and reaction time. Nailing micro-expressions and turn-taking fixes the exact friction points that make people hang up on AI callers.

The catch will be operational cost and edge latency. Orchestrating three specialized models alongside real-time web search and vision processing requires serious compute overhead. While Tavus envisions digital humans becoming as ubiquitous as smartphones, teams will need to see whether the per-minute API economics make sense outside of high-ticket sales and enterprise support.