Introducing Phoenix-4, the most advanced real-time human rendering model ever built. Its the first real-time model to render every pixel at runtime, generate and control emotional states, listen actively, and actually behave the way humans do in conversation.

Humans evolved to communicate face to face. We convey meaning through reactions, emotions, and nonverbal signals that tell us someone is engaged, understands us, or cares. To understand this dance, Phoenix-4 learned how humans actually move in conversation from thousands of hours of real data. Expressions aren't pre-programmed, they emerge naturally- just like in real conversation.

As result, Phoenix-4 is a real-time human behavior engine, not just a video generation model, unlocking natural conversations. The model: - Expresses emotion naturally with continuous, emergent microexpressions - Actively listens with context-aware head pose, gaze, and nods - Preserves identity across subtle and extreme expressions - Generates full-face behavior (eyes, head, hair, lips) in real time - Runs at 40fps at 1080p with low latency

Phoenix-4 is the first model to be able to express a full range of human emotion, generating and controlling 10+ emotional states, including happiness, sadness, anger, surprise, fear, and excitement, while transitioning between them seamlessly in real-time. You can direct emotion through your LLM or let the model respond contextually through Raven-1 perception.

Listening is the other half of every conversation. Phoenix-4 is the first of its kind to actively generate contextually correct listening frames, rather than just looping over a pre-recorded video. It is a full-duplex model, which means it hears, understands, and reacts while the user is still speaking. Context-aware nods, gaze shifts, and micro-expressions then respond to what's actually being said.

Phoenix-4 is built on a novel hybrid Gaussian–diffusion architecture that generates every pixel in real-time. Every frame is new. Rather than animating predefined states, the model learns a continuous behavioral representation of human movement and reaction. It generates full-face and head pose behavior, synchronized to the rhythm and emotion of the conversation. All in real time at 40fps in 1080p.

Healthcare. Therapy. Education. Any interaction where empathy matters. When an avatar can actually respond to how someone is feeling, conversation becomes something closer to connection. Phoenix-4 sets a new standard, bringing human-AI interaction closer to human than ever.

Read the full research behind Phoenix-4, including the architecture, training approach, and how emotion control works end to end in our latest blog.

Phoenix-4 is live today in the Tavus Platform, APIs, and PALs with stock and custom replicas available now. Try the demo and feel the difference when a machine finally learns to be human.

At this rate, people will start creating avatars to send them to interviews for new jobs, do podcasts, and speak online. How are we going to differentiate real from AI?

they already are :) check out for yourself at

@svpino Do you have a github?

@svpino Check out our docs here:

Seriously, thats really impressive!

you know how much this means to us... this much. ✋-------------------------------------------- ✋

What interests me here is the listening state. In human interaction, credibility is built while the other person speaks. If AI can manage reaction, timing, and emotional shifts coherently, the perception of presence changes significantly. @tavus partner

perfect for scammers

@r39ister the goal is never to deepfake, and this is why we have specific guardrails in place to avoid this. ❤️

@r39ister like what? im interested in that

@tavus Таvus, see? I told ya. 🥸

This is just amazing 🤯🤯🤯🤯🤯 - I just tried out the application and mind blown 🤯

You know.....it is sad that when you look at job postings, it is all engineering. These products will need perspectives from liberal arts. I took time to say this because your product has set up a cottage in the uncanny valley. Very clever approach, and execution, but you need to hire some human computer interaction and counseling psychologists to get out of that valley. Great progress in terms of winning the turing test.

we agree that getting out of the valley takes more than engineering, that's why we trained on real human behavior instead of programming it.

Consider opening research partnerships and internship opportunities with university psychology and counseling programs. They could design qualitative and quantitative experiments to complement your data-driven approach. My point is that the real edge comes from integrating these human-centric disciplines, rather than relying solely on raw data. Great win!

scammers rn:

our identity preservation tech has something to say to scammers 🤜

how?

I think the technology is genuinely impressive, and that is also exactly what makes it worth discussing seriously. When an AI can render emotions in real time and behave the way humans do in conversation, the line between real and synthetic interaction starts to blur. I am not opposed to the tech, but I think people don't realise how much of their privacy they are trading when they engage with systems like this. We are building a world where you can't trust what you see. That has real consequences beyond the product itself.

the guardrails we're building are just as important to us to focus on as we continue to build out the future of human computing. ❤️

This is really impressive. The face expressions and emotions look so natural. Great work 🔥

this is what happens when you train on real human conversation instead of programming expressions by hand. 🤚❤️

Kudos to the team. Real human rendering model getting real!!!

thank you, Arsh!!

🐦🔥

Most of communication is nonverbal. To this point, it's always been the blind spot. Wild to finally watch someone actually go after it. Nice work.

thank you, also please try the demo and let us know what you think 😊 demo:

This is the first time an AI avatar actually feels like it’s listening instead of waiting to talk. Huge unlock for real conversational UX.

listening is half of every conversation, it deserved a model that treated it that way.

Great work from the team!

❤️

You have never met an AI like this before! Congrats to the @tavus crew for shipping an incredible, SOTA rendering model!

If Phoenix 4 is building on top of real time pixel generation while preserving identity, that’s not just an incremental update, that’s a serious leap in generative video infrastructure.

incremental was never really our style. 🤭

This is a huge step forward! When models stop just generating output and start responding like humans, the interface itself changes. Congrats on the launch, wild to watch this shift happen in real time.

that shift from output to behavior is the entire mission. 🫡

This feels like a real inflection point for human-AI interaction.

👆👆👆

Adding EQ is a total game changer. Congrats on the launch. It’s a wild time to be living on this planet 🔥🔥

everybody in AI has been so worried about IQ that we completely forgot about EQ.

Generating every pixel at runtime while preserving identity is seriously impressive engineering

wait until you see the Phoenix-4 demo:

Congrats! 👏 Building a grief AI companion with @elevenlabs + Claude Sonnet 4.5 and exploring video avatars. Big week for real-time AI with @tavus & @Anam__ai making waves. What should developers know about what makes you each similar/unique? 🦾

@elevenlabs @Anam__ai happy to let builders compare for themselves. Try the demo, the difference is pretty immediate! ☺️ demo:

tavus cooked with this one! 70% of human communication is nonverbal. Every avatar model before this ignored that entirely. Phoenix-4 is the first one that actually took it seriously.

spot on! are you building with Tavus yet?

its getting better every update!!! 👀🎉

more research, better models, infinite use cases.

Incredible!

Based model and RT rendering approach. Dope for AR/VR apps. Created you a spotlight on our platform guys, let's goo!

🔥 I’m actually emotional right now, that's absolutely insane!

so are we, max... so are we. 😭

Very cool milestone for real-time human interaction. Congrats on the launch✨ @tavus

They talk, listen, nod and actually feel alive.. its already too real wow

and... they are even fully agentic. 🤖

feels like a real leap. healthcare and therapy use cases are going to be huge for this congrats team!

spot on! go demo Phoenix-4 yourself here:

👀

@IbnMaghrebi @AvdullahYousef @AlboranNomad

From pre-programmed expressions to emergent, context-aware behavior, this feels like the future of human-AI interaction. Very impressive.

This is epic, and pretty honored you named it after my home town. Excited to try it out!

Phoenix-4 🤝 Phoenix, AZ let us know what you think!

this is super awesome. looking into the documentation now, perhaps worth experimenting!