Avaturn Live
AI avatars that react to you in real time. Now OPEN SOURCE.
Start new thread

AVTR-1 Real-Time Open Weights Model - Generating uncanny AI avatars is now open source

by•
The best real-time avatar model in the world is now open source with open weights. Take the model, tweak it, and use it at $0 cost. What's unique: our model listens while you speak — full-duplex; the avatar reacts in real-time, with minimal latency. • Every frame is generated, avoiding annoying animation loops from pre-rendered playback. • Full streaming infrastructure included so you can get started right away.

Add a comment

Replies

Best
Maker

Hey Product Hunt 👋

I'm Sergei Sherman, CEO of .

Today we're releasing AVTR-1 — an open-weights real-time AI avatar model that sets a new state of the art on key benchmarks.

If you're building anything with real-time AI avatars, AVTR-1 is for you.

✍️ Here's what makes AVTR-1 different:

  1. The whole face is generated. Not just the lips swapped onto a pre-recorded clip. Every pixel of the avatar's face, top of the head to the chin, is generated in real time, frame by frame.

  2. Native duplex — the avatar actively listens. The model is generating all the time, whether the avatar is speaking or listening. Just like a human on a call, the avatar's face responds to your words and your tone in real time. The brow lifts at word three because you sounded surprised at word three, not after the sentence ends.

For three years, "real-time avatars" have meant pre-recorded video with a generated mouth pasted on top. We threw out the recording.

🎯 Why you want AVTR-1:

  • Open weights. Free for personal, research, and any commercial use under $10M in annual revenue. Commercial licensing above that, through us.

  • Sub-200ms end-to-end on one A100 or 4060. Runs on youd device in a data center, in the cloud.

  • Avaturn Streamer included — the open infrastructure layer for real-time avatars. Accepts AVTR-1 or any other open-weight real-time video model as a drop-in. Plug in your video model on one side, your conversation backend on the other.

  • Reference avatars out of the box. Model cards, license-cleared, deployable today.

  • Launch-partner examples in the repo with Cartesia and Pipecat on day one.

🏗️ One thing we're explicitly NOT launching — and want the industry to build with us:

A public, vendor-neutral leaderboard for real-time AI avatars. The category needs a transparent scoreboard, one the ecosystem runs together. Clear, public competition is the only way improvement happens fast.
We're inviting every other vendor, every open-source contributor, every researcher to help us build it.

🎉 Everything is live today:

Real-time generated video is the next frontier. Every previous wave — text, then real-time audio — produced an open layer the category built on. We're shipping that layer today: model and orchestration both.

Drop questions, feedback, or what you're building below — I'll be here all day 🚀

— Sergei

the active listening part is what separates this from every other avatar tool. current ones just talk at you with dead eyes while you wait. if the expression matching actually works in real time this changes how you build AI sales and onboarding flows

Maker

 Thanks Tina! Yes, this is our differentiator, thanks for commenting on this!

Hunter

Tried the demo before launch — the lip sync is noticeably better than what I’ve seen with other generative avatars. But walked me through something deeper: active listening coupled with empathetic response.

You know how you can kind of say anything to most AI avatars (e.g. “my mom died”, or "omg there's a murderer outside my window!") and they'll just blink, cycle their idle loop, nod and say something like “oh, that’s nice to hear.” These bots are just mouths on a timer with zero semantic read.

AVTR-1 generates every pixel of the face in real time, frame by frame. When meaning shifts in what you’re saying, the expression shifts to match — e.g. brow lifting at word three because the content warranted it, not just because the sentence ended.

For developers: there’s no Pipecat equivalent for video agents right now. is shipping the full stack — model weights, streamer, sync layer, reference avatars. Bring your own GPU, and you're ready in 15 minutes. Open weights is a big deal, and it's all free if your business is under $10M ARR.

Maker

 Thanks Chris, we rreally are excited on giving this to developers and end users. Theoretcially? You can now generate enldess avatar content at $0. Exciting times!)

Maker

Hey guys! We are so excited to show you our new model: avatars became even more realistic and reactive. The awesome thing is that active listening is now at another level: avatars are reacting to your speech like a real person. If you are not a technical person like me, you can simply go to our website and talk to our avatars to see how cool that is! If you are a developer yourself, check out our github: we opened our model!

 I just spoke with Ben. When you say it’s open source, can we create other characters, or do we have to use the ones you already have?

Maker

 Yes, you can create your characters

the 9x faster claim is interesting but faster than what exactly. the baseline matters a lot here. sub-300ms felt like magic two years ago and now it's table stakes for anything calling itself real-time. curious what the actual latency numbers look like end to end, not just the generation side

Maker

 
On 300ms: it's not generation time. Generation actually takes 80-90ms depending on GPU model. The 300ms is the model-side pipeline latency that comes from audio context buffering — you need to look ahead at a chunk of audio before producing a lip-synced frame, which is fundamental for any audio-driven avatar. Doesn't include the latency of wherever the avatar's speech is coming from (STT-LLM-TTS pipelines or speech-to-speech models) or the time required to deliver frames over the network to the viewer. Add those in and true E2E depends on the setup, location, and network conditions. Our results match or exceed proprietary competitors, and we're open sourcing the model and code anyway. Try the online demo and judge for yourself. On 9x — that was against an offline non-realtime generator, so it's more "real-time vs not" than a clean like-for-like baseline.

How is it different from Tavus or HeyGen?

Maker

 are you using Tavus or Heygen? Like daily? Speaking with them for hours? I guess not. The new model is trying to achieve this. Better avatars, that you can speak naturally, and by open sourcing it - you can install it on your device at $0 , which is attractive price I guess. But more importantly, we open source so anyone can contribute so this model moves faster on development towards inflection point where you will feel that avatars are just like human

Tried the demo, looks impressive. Why have you decided to open source your model?

Maker

 Thanks! The reason is simple, open source moves faster, and growing the entire market wiht more developers coming in. We want to move fast and we believe the market can be 10X bigger then it is today, but for that you need also a good model. As soon as we saw the benchmark results and how user engage with it we realized this is the first model that can deliver this promise, so we want to spread it as fast as possible. Note that we also released our streamer so our competitors can use it for better streaming, the entire industry needs to jump over a certain quality bar to be good enough - > We are helping this happen today.

What GPU is needed?

Maker

 Multiple GPUs supported here is example from Github repo

GPU

Latency / 5-frame chunk

Real-time factor

L40

84 ms

2.4×

A100

91 ms

2.2×

RTX 4060 Ti

166 ms

1.2×

RTX 3070

181 ms

1.1×

L4

202 ms

0.99×

RTX 3060 Ti

206 ms

0.97×

RTX 4060

232 ms

0.86×

Love this! I am working on a project to enable voice as primary interface for agents doing real work would love to integrate with AVATURN! Tried using with Tavus before but too pricey, how can we scale this to multiple users any tips on the same? Love the project and initiative, congrats on the launch!

Maker

 You will need some code to manage stream lifecycles. We plan to release it as well later. And, you're going to need some autoscaling configuration and a fleet of GPUs running with some over provisioned capacity, so that your users will get the avatars in a reasonable time.

Or you can just use our managed API already available.

Really great to see truly open models for realtime video (and a nice technical overview post). Congratulations on the launch, team!

Maker

 Thanks! Very happy about our collaboration with Pipecat, Github repo and Avaturn.live API both are happily running on it:)

12
Next
Last