# Tavus launches Phoenix-4.5 with upper-body motion

> Phoenix-4.5 adds upper-body movement and produces video 134 milliseconds after audio, Tavus said. The model is now available for all PAL avatars through the company’s API and no-code PAL Maker.

- Publication: Go Big News
- By Nifal Adam, Go Big News Staff (Staff News)
- Category: AI
- Published: 2026-09-11T12:46:01.023+00:00
- Canonical: https://www.gobignews.com/ai/tavus-launches-phoenix-4-5-with-upper-body-motion

---

[Tavus](https://tavus.io) announced Phoenix-4.5 on Sept. 10 as a real-time human rendering model for the AI avatars it calls PALs. The company said the model converts audio to video in 134 milliseconds, which it said is 25% faster than any other model on the market.

In posts on X, Tavus said Phoenix-4.5 adds movement across the upper body and offers richer facial animation, more expressive emotion and micro-movements. It also reported a 16% lip-sync improvement over Phoenix-4 and said 64% of faces that did not train on the prior model now train successfully.

A Sept. 10 research post by Minh Anh Nguyễn and Haiyao Xaio described a change in how the system generates video. According to the post, Phoenix-4 rendered only a small region around the face, about 512 by 512 pixels, then composited it onto recorded body footage. Phoenix-4.5 instead creates the frame with the face, head, neck, shoulders, torso, clothing and surroundings generated together through a half-body network.

The new model is Tavus’ first zero-shot Phoenix model, the company said. It said a preview is available in about a minute, while fine-tuning continues in the background and takes about two hours, half the time required for Phoenix-4. Tavus also said long hair, glasses, earrings and headbands train more reliably than before.

Tavus said it upgraded all of its existing stock faces to Phoenix-4.5 and is adding 50 new ones. Its documentation says the model supports chest-up or waist-up portraits, rather than full-body framing.

On an internal paired test set, Tavus reported an AVSR lip-sync score of 0.304 for Phoenix-4.5, compared with 0.261 for Phoenix-4, while identity similarity remained roughly unchanged. The company said the stack runs at about 35 to 40 frames per second end to end on slower machines.

Tavus, founded in 2020 by CEO Hassaan Raza and COO Quinn Favret, is headquartered in San Francisco. It offers a developer API for real-time PAL conversations, a no-code PAL Maker and managed enterprise deployments.

---

## Companies in this story

- Tavus: https://www.gobignews.com/company/tavusio

---

## X profiles in this story

- @tavus: https://www.gobignews.com/profile/tavus