Meta says anyone in the United States can now call and video chat with Muse, its consumer AI agent, through a live animated avatar. TechCrunch reported the rollout on November 5, 2026, citing Meta's announcement. The feature is called Muse Voice. It runs on mobile, web and Mac, and Meta says it will later come to its AI Glasses and the Muse Charm handheld device. The underlying models were first shown in September, so this is a rollout story, not a new model story.
This post covers what shipped, what Meta has measured and published about the avatar system, what the figures do and do not show, and what to watch next. For wider Muse context, see our posts on Muse Voice Transcribe, the Muse Charm style device and Waddle World, and the Muse privacy questions raised by TIME.
TL;DR: what changed and what did not
| Question | Answer | Source |
|---|---|---|
| What launched? | Muse Voice: call and video chat with Muse as an animated avatar | TechCrunch |
| Who gets it? | Everyone in the US | TechCrunch |
| Which devices? | Mobile, web and Mac. Glasses and Muse Charm later | TechCrunch |
| Are the models new? | No. Introduced September 23 | Meta AI Research |
| Latency claim? | About 870 ms to first synchronized byte | Meta AI Research |
| Independent test? | None found | Our search |
| Price? | Not stated in the material we read | Unknown |
| Other countries? | Not stated | Unknown |
What Meta shipped on November 5
According to TechCrunch's Sarah Perez, Muse Voice lets users call and video chat with Muse. The avatar is animated, nicknamed Jolly, and users can customize it into something that looks like a person, an animal or another character. While the avatar speaks, Meta says it uses "subtle expressions and gestures" that match the tone of the conversation.
Wang announced the rollout on X and in a blog post, and shared a demo in which his own Muse appears as a fluffy white dog. In the clip, as TechCrunch describes it, Wang jokes about posting memes on X, and the dog suggests adding disclaimers to memes that might be read as Meta endorsements. We could not load the Meta blog post itself, so the availability details here rest on TechCrunch's report. Check Meta's newsroom for the exact wording.
TechCrunch also notes the timing: the rollout lands before Meta's planned launch of a Tamagotchi-like wearable for Muse, due before year end. When that device arrives, the avatar will appear on its small screen. Our Muse Charm coverage explains what is known about that product line.
How the avatar works, from Meta's research post
Two speech bubbles passing a bead, showing how Meta Muse Voice keeps turn-taking in a live avatar chat
Meta published the technical account on September 23, 2026 in Bringing Your Muse to Life. The core claims:
- One stream for voice and face. Muse Realtime Voice produces a stream of speech tokens that carry both what is said and how it is delivered. An audio decoder turns the tokens into speech, and Muse Realtime Avatar reads the same stream to generate the matching visuals. Meta says sharing the stream keeps voice, lip motion and expression in sync.
- A diffusion transformer. The avatar model is audio-driven, conditioned on the speech tokens, reference media and a rolling window of recent video latents. It generates video in short causal chunks, and each finished chunk becomes motion context for the next.
- Any image becomes a character. Given reference media, the model can animate a photographic portrait, a full-body illustration, an animal or an everyday object, with facial, hand and full-body motion.
- Distillation. A bidirectional teacher uses 40 diffusion steps with three-way classifier-free guidance, which is 120 model evaluations per chunk. Meta distills this into an unguided two-step student, a 60x reduction in evaluations, using self-forcing and distribution matching distillation.
- Serving. The system streams 448x768 portrait video at 25 frames per second with about 870 ms of latency, measured from the end of the user's turn to the first byte of the synchronized response. On one GB200, each step produces eight frames (320 ms of playback) in 20 ms. Meta says serving optimizations raised capacity 8x over a two-step BF16 baseline, enough for 12 concurrent video sessions per GB200.
The engineering details are plausible and specific, but all of them are Meta's own measurements. Readers who want to judge them can compare with other real-time video work we have covered, such as the Atmanity real-time avatar claims, which we also checked claim by claim.
What the rater study does and does not show
Meta compared Muse Realtime Avatar with Runway Characters and HeyGen LiveAvatar, which it calls the two leading commercial avatar systems. Raters held two- to three-minute conversations with each system, using each product's native live-call experience and matched avatar identities. They compared visual quality, synchronization, character consistency, mannerisms and other dimensions. Meta says raters preferred Muse Realtime Avatar overall and in every dimension, with the mannerism comparison against Runway Characters not statistically distinguishable from parity.
Three cautions apply:
- It is a company-run study. The rater pool, sample size and prompts are not described in the text we read.
- Matched identities are not the same as the shipped experience. Meta's own note says the post's examples show model capability and do not all reflect avatars available in the Muse app.
- Preference is not usefulness. A more pleasant face does not mean Muse gets more tasks right. TechCrunch's reading is that Meta thinks anthropomorphizing the agent could make consumers feel more connected to it.
Why Meta wants a face on its agent
Muse competes with chat assistants that already offer voice. Voice alone is a feature. A persistent, customizable character is a product that users can attach to. Meta has already leaned on character and personality in Muse, as we discussed in what Meta got right with Muse.
The business backdrop matters. Our earlier checks of Meta's usage claims found a mixed picture: we looked at the reported plateau in daily users and the 73% day-one retention claim. A more engaging interface is one lever to move retention. Whether it works is an empirical question that Meta has not published data on.
Safety and privacy points to know
Soft waveform with a green peak, representing the voice stream behind Meta Muse Voice avatars
Meta's research post lists its safeguards for the avatar system:
- Watermarking. Muse Realtime Avatar uses Meta Video Seal to embed a durable, invisible watermark throughout generated video, without adding latency.
- Age limit. Muse is for users aged 18 and over.
- Ongoing work. Meta says it will keep strengthening protections as embodied AI evolves.
Questions the post does not answer, at least in the text we read:
- Whether the avatar can be built from a photo of another person, and what checks exist for consent.
- How long video or voice from calls is stored, and whether it is used for training.
- How the camera feed is handled during a video chat, if the user turns it on.
These matter more because TIME reported earlier that Muse can build a profile of a user and their contacts. We cover that in Meta Muse builds a dossier on you and your contacts. Before you start a video chat, read the in-app privacy settings and check whether chats are used to improve models.
How to try it and what to check first
If you are in the US, update the Muse app or open Muse on the web or Mac. Look for the call or video option. A sensible first session:
- Start with voice only. Check whether the app asks for microphone and camera permissions separately.
- Open the avatar settings and customize the character. Note that the research post shows that images can be animated, but the app may offer fewer options than the research examples.
- Ask a task that uses a connector, such as the ones in Meta's developer connectors platform, and see whether the avatar changes anything about how the agent behaves.
- Review the chat history and delete anything you do not want stored.
If you build voice agents, the more useful takeaway is architectural. Sharing one speech-token stream between the voice model and the video model is a clean way to keep sync without a separate alignment step. The 60x step reduction through distillation is also a pattern worth studying for any real-time generative system.
What is still unknown
- Pricing and limits. We found no statement on whether video minutes are capped or later paid.
- International rollout. TechCrunch reports US availability only.
- Independent latency. Meta's 870 ms is measured in its own stack on GB200 hardware. Real-world latency depends on network and device.
- Impact on retention and trust. No data yet.
Bottom line
The Muse avatar is not a new technology on November 5. It is the first time most US users can try it. The research post gives a detailed, credible technical account, and the safeguards it names are sensible. The proof will be in independent tests of latency and in whether users find a talking character more useful than plain voice. Until then, treat the 870 ms figure and the rater study as Meta's claims, and read the privacy settings before you turn the camera on.
Related reading
- Muse Voice Transcribe: Meta's real-time speech-to-text
- Meta Waddle World and the Muse Charm style device
- Meta Muse builds a dossier on you and your contacts
- What Meta got right with Muse
- Meta Muse daily users plateau: what it means
- Meta opens Muse to third-party developer connectors
- Meta Muse comes to Windows
Details are accurate as of November 5, 2026. Product availability and features may change.