Video-calling an AI: Meta's Muse gets an avatar with facial expressions and gestures

What's going on

Meta has given its personal AI agent Muse a voice and a face. Since November 5, 2026, users in the US can call their Muse or video-chat with it. An animated avatar appears on screen and reacts to the conversation: it changes its expression, moves its lips in sync and gestures as it talks. The feature works in the Muse app on smartphones, on the web and on the Mac. Meta's Chief AI Officer Alexandr Wang announced it in a blog post and on X. Meta first showed the technology in September at its Connect 2026 developer conference.

According to Android Authority, Muse has only been on the market for about two months and has already picked up two new models called "Muse Wonder". So far the service is only available in a handful of countries, including the US, Canada, Japan and Singapore. The new voice and video feature is US-only for now. According to shiftdelete.net, Meta hasn't given a timeline for other countries. Whether and when Muse will come to Europe is unclear.

How a call with the AI works

You start a conversation by tapping the phone icon in the top right of the Muse chat window. Then you just talk instead of typing a carefully worded prompt. You can interrupt Muse mid-answer. According to shiftdelete.net, the conversation even keeps going while Muse handles tasks in the background. You can choose from several hundred voices.

One important point: the AI can't see you. Unlike a regular video call, Muse doesn't use your camera to analyze your face. The avatar's reactions are based only on your voice and the flow of the conversation. It visibly "listens" and starts reacting while you're still talking.

You can pick from three views:

• voice only, like a classic phone call,
• avatar next to the chat,
• full screen for a "face-to-face" conversation.

Out of the box, the avatar is called "Jolly", a cute animated character. But you can redesign it however you like: as a different person, an animal, your own character or even an everyday object. According to Meta, you can create characters from reference images or videos. Their look and movements are supposed to stay consistent throughout the conversation. In his demo video, Wang turned his Muse into a fluffy white dog and joked with it about whether his memes might get him in trouble at work.

The tech behind it: from 120 compute steps to 2

According to Meta, as reported by the Korean outlet AI Times, two models work together in a single continuous data stream:

1. Muse Realtime Voice generates speech tokens. These are small chunks of data that capture what is being said and how, meaning tone and emphasis.
2. Muse Realtime Avatar turns these tokens into picture and sound using a speech decoder and an avatar model. That keeps voice, lip movement and facial expressions in sync.

The video itself comes from a model based on a diffusion transformer. Diffusion models build an image out of noise and remove that noise step by step until the final image emerges. Normally that takes a lot of compute passes. So Meta uses distillation: a fast "student" model learns to imitate the output of a slow "teacher" model. According to Meta, it now needs just 2 steps instead of 120.

Meta measures the roughly 870-millisecond delay from the moment you stop talking until the synchronized audio and video response begins. Just under a second is noticeable in a conversation, but still tolerable. Meta worked with Nvidia on the server infrastructure. A single GB200 GPU is said to handle 12 video sessions in parallel, and overall capacity is eight times that of the original base model. These numbers come from Meta itself. There are no independent measurements yet.

Safeguards: watermarks and an age limit

Meta embeds the invisible watermark Video Seal in every generated video. It's a kind of digital stamp you can't see with the naked eye, but machines can read it. It's meant to show that a video was made by AI and make misuse harder, for example if someone passes off recordings as real footage. According to Meta, the watermark doesn't slow down real-time interaction. Muse is also only available to people aged 18 and older.

What's next: Muse Charm and AI glasses

The launch also lays the groundwork for new hardware. Before the end of the year, Meta plans to release the Muse Charm, a palm-sized device that India Today compares to a Tamagotchi. Your Muse avatar is supposed to "live" on its small screen. That means you could talk to the AI and give it tasks without pulling out your phone. Real-time conversations are also set to come to Meta's AI glasses over time.

Our take

ChatGPT, Gemini and others have had voice modes for a while. That's why shiftdelete.net calls real-time voice "table stakes" for AI agents. Meta is betting that a responsive face, plus tasks running quietly in the background, will make Muse feel more like a companion than a chatbot. Technically, delivering smooth video with under a second of latency is impressive. How well it holds up outside hand-picked demos remains to be seen. And for most people outside the US, this is just a preview for now.

There are critical voices, too. Analytics Insight warns that people talking out loud to a friendly face tend to reveal more about their health, relationships or private lives than they would when typing. Especially with a company like Meta, which makes its money from data, that raises the question of how these conversations are stored and used. Shiftdelete.net also flags the default privacy settings as an open question. Meta doesn't give specifics on this in the available sources. Usage limits, possible costs and the timeline for other countries are also unclear.