Muse Realtime Avatar vs Muse Video: Two Meta Video Products, One of Which You Can Actually Use

Search either name and you will find pages that treat them as siblings, which is understandable and wrong in a way that costs people time. Muse Realtime Avatar is Meta's embodiment layer — a live renderer that turns the speech-token stream from Muse Realtime Voice into a talking, gesturing character — and since 5 November 2026 it has been rolling out inside the Muse app in the United States. Muse Video is Meta's prompt-to-video generator, previewed on 7 July 2026 alongside Muse Image and never released: no API, no price, no resolution, no clip length, no date. One of these is a product in an app. The other is a demonstration with a name and a leaderboard position. The reasoning half of the Muse family is reachable on one API, 0% markup while the rendering half has no endpoint at all.

So the honest comparison is not "which generates better video." It is "which of these two things is the thing I am looking for," because the answer decides whether you have any decision to make at all. For almost everyone arriving at this page, the answer is Muse Realtime Avatar, and the reason is not that it is better — it is that Muse Video is not available to compare against.

Two names, two very different states of existence

The dates do most of the work here, and they are worth laying out plainly because the two products are five months apart and only one of them moved.

Muse Video was announced on 7 July 2026 in a post titled "Introducing Muse Image and Muse Video," published under the Meta Superintelligence Labs byline. That post's opening line draws the distinction Meta has never revisited: "We're excited to launch Muse Image and preview Muse Video." Muse Image shipped that day across the Meta AI app, meta.ai, Instagram Stories in the US and WhatsApp in limited countries. Muse Video did not. The availability line, repeated twice on the same page, is "Muse Video is coming soon to creators and Meta AI."

Muse Realtime Avatar was announced on 23 September 2026 at Meta Connect, in a research post called "Bringing Your Muse to Life," and was described at the time the way Muse Video still is — a capability demonstration with no API, no price and no date. Then on 5 November 2026 Meta posted that it was "rolling it out in the Muse app in the US, giving people a more natural way to engage with their personal agents."

That is the whole timeline. Muse Video has been "coming soon" for four months. Muse Realtime Avatar went from research post to consumer rollout in six weeks. If you are trying to work out which Meta video product is further along, the answer is not close, and it is the opposite of what the naming suggests: "Muse Video" sounds like the flagship video product, and "Muse Realtime Avatar" sounds like a feature.

What each one is actually for

Set the shipping status aside for a moment. The two systems solve different problems, and the difference is architectural rather than a matter of quality.

Muse Video is a batch generator conditioned on a text prompt. You describe a scene, it produces a clip. Meta's claims for it are short and relative — "competitive performance in prompt adherence, visual fidelity, and temporal consistency" — with no named comparison set, and the company names its own gaps in the same paragraph: "audio-video synchronization and physically accurate fast motion." It generates sound natively rather than adding it afterwards, and it is built on the same pretraining base as Muse Image.

Muse Realtime Avatar is a live renderer conditioned on speech. It does not decide what happens in a scene; it renders a character you supply while a conversation happens. Meta's description is an "audio-driven, Diffusion Transformer conditioned on speech-token stream, reference media, and a rolling window of recent video latents," generating "video in short causal chunks." Its input is not a prompt. Its input is the token stream coming out of Muse Realtime Voice, which carries both what is being said and how it is being delivered.

Put bluntly: Muse Video would generate a video of a person talking if you wrote a good enough prompt. Muse Realtime Avatar is the person talking, responding to you, for as long as you keep talking. The first is a rendering job with an end state; the second is a live system with no end state until the call ends.

The contrast that decides real use cases

Here is the same comparison on the dimensions that matter, with every figure attributed to whoever published it. Meta has published numbers for Muse Realtime Avatar; for Muse Video it has published almost none.

• Availability — Muse Realtime Avatar is live in the Muse app in the US, 18+, no API; Muse Video is a preview with no general availability, no waitlist and no announced date.

• Conditioning — Muse Realtime Avatar consumes a speech-token stream plus reference media plus recent video latents; Muse Video takes a text prompt and, per Meta, prioritises prompt adherence and temporal consistency.

• Output envelope — Muse Realtime Avatar streams 448×768 portrait at 25 fps; Meta has published no resolution, frame rate or clip length for Muse Video at all.

• Duration — Muse Realtime Avatar runs for the length of a conversation, generating in causal chunks with a rolling latent window; Muse Video's clip length is undisclosed.

• Latency — Muse Realtime Avatar is reported by Meta at approximately 870 ms from end of turn to first byte of synchronized voice and video; Muse Video has no published latency, because there is no interactive path to measure.

• Audio — Muse Realtime Avatar's audio comes from the voice layer's decoder, sharing the token stream with the picture; Muse Video generates audio natively, and Meta lists audio-video synchronization as a current weakness.

• Pricing — Muse Realtime Avatar has no price and is bundled into Muse subscriptions at $20 or $100 a month; Muse Video has never had a price published.

• Watermarking — Muse Realtime Avatar uses Meta Video Seal, which Meta says embeds a durable invisible watermark without adding latency; for Muse Video, Meta says only that it plans "to extend Content Seal to video soon."

Why the shared token stream matters more than it sounds

The single most useful thing to understand about Muse Realtime Avatar is that it does not do lip sync. It never has, and that is the design.

Most avatar systems work in two passes: generate or record the audio, then run a separate model that makes a face match it. The seam between those passes is why so many avatars look dubbed — the mouth is correct and slightly late, the expression is plausible and unrelated to what is being said. Meta's approach removes the seam by construction. Muse Realtime Voice emits speech tokens carrying content and prosody together; an audio decoder turns them into sound, and the avatar consumes the same tokens to generate the picture. Voice, lip motion and expression stay locked because they come from one stream.

The second half of the design is the rolling latent window. A chunk-based video generator drifts: small errors compound, and by minute three the character has quietly become someone else. Meta feeds each completed chunk's newest latents forward as motion context for the next, which carries appearance and mannerisms across turns while keeping per-chunk compute bounded. Meta's own framing is that this is what allows generation "to continue for as long as the conversation does."

Muse Video has no equivalent problem to solve, because it has no conversation. Its stated priorities — prompt adherence, visual fidelity, temporal consistency — are the priorities of a clip that will be watched once and judged as a clip. Different objective, different architecture, and no meaningful way to rank one against the other on a single axis.

The numbers Meta published for the avatar, and the ones it did not

Muse Realtime Avatar arrived with an unusual amount of measurement for a consumer feature, and the discipline is to keep each figure attached to what it measures.

The 870 ms is a first-byte figure — end of your turn to the first byte of the reply — not time to a complete answer and not steady-state latency ten minutes in. The 448×768 at 25 fps is a portrait frame at a cinematic frame rate, not a landscape delivery format. The serving claim is that an optimized stack, including four-bit quantization-aware training and persistent KV caches built with NVIDIA, raises capacity 8× over Meta's two-step BF16 baseline and enables 12 concurrent real-time sessions on a single GB200, at 2.5 ms of model time per frame. The distillation claim is 60× fewer model evaluations per chunk — 120 down to 2 — which is a per-chunk compute reduction, not a sixty-fold improvement in what a user feels, and Meta's own chart shows human preference moving from 45% to 55% rather than holding flat.

The evaluation is the part worth reading closely. Meta compared against Runway Characters and HeyGen LiveAvatar using matched avatar identities, with raters holding two- to three-minute live conversations and then comparing visual quality, synchronization, character consistency and mannerisms. Meta reports being preferred 78% to 22% and 88% to 12% respectively, and discloses in the same post that the mannerism comparison against Runway Characters was not statistically distinguishable from parity. All of it is vendor-run, on vendor-chosen raters, in conversations the length of a coffee order.

What Meta has not published is just as important. There is no independent benchmark of Muse Realtime Avatar anywhere, and the caveat that its examples "do not all reflect avatars available in the Muse app" has not been replaced by a description of what the app build supports. Whether you can supply your own reference image, whether every subject category in the research post is enabled, whether sessions have a time limit, whether the 448×768 envelope is what a phone actually renders — none of that is documented.

Muse Video's number problem is worse in kind. The only quantitative claim Meta has made is a leaderboard position: "On Arena, Muse Video ranks No. 3 in human-preference Elo for text-to-video at the time of writing," with the underlying chart captioned "Arena Elo rankings as of July 5, 2026." That is a vendor-reported, human-preference, time-bounded figure from two days before the post was published, and it has almost certainly moved since. There is no benchmark, no resolution, no clip length and no price to pair it with.

What the Muse family actually ships today

This is the part that gets lost in the model-by-model coverage, and it is the part a reader planning anything needs. Meta's Muse line is not one product with variants; it is a set of things at wildly different stages, and only some of them are callable.

• Muse Realtime Avatar — shipping to consumers in the Muse app in the US since 5 November 2026, no developer surface.

• Muse Realtime Voice — the conversational layer underneath it, also without a public API.

• Muse Video — previewed 7 July 2026, never released, no date.

• Muse Image — shipped 7 July 2026 across the Meta AI app, meta.ai, Instagram Stories in the US and WhatsApp in limited countries.

• Muse Voice Transcribe — a real, priced speech-to-text endpoint, documented as speech-to-text only with no speech-to-speech conversation surface.

• Meta Muse Spark 1.2 — the reasoning model the agent runs on, and the one Muse checkpoint that is routable on OrcaRouter today, as meta/muse-spark-1.2 at $1.25 per million input tokens and $4.25 per million output over a one-million-token context that accepts text, images, video, audio and files. Its predecessor, meta/muse-spark-1.1, sits beside it on the same list.

We do not carry Muse Realtime Avatar, Muse Realtime Voice or Muse Video. That is not a hedge — the live catalogue returns exactly two Muse entries, both Spark checkpoints, and saying anything softer would imply we route something we do not.

The question this comparison actually answers

If you arrived wanting to know which Meta video model to use for a project, the answer is Muse Realtime Avatar, and only if your project is a conversation with a character inside the Muse app. Everything else — clips, scenes, anything with a delivery file — has no shipping Meta option at all, which is why teams doing that work are on MiniMax H3, Kling 3.0 or HeyGen Video 1 rather than waiting.

If you arrived because the names are confusing, the short version is that "Video" is the one that has not shipped and "Realtime Avatar" is the one that has. That inversion is not Meta being clever; it is what happens when a research preview and a consumer feature are announced five months apart under a shared family name.

And if you are building something that needs both halves — an agent that reasons and an agent that appears — the split is unavoidable today. The reasoning half is available and priced. The appearance half is a consumer feature with no door.

What would make this comparison worth writing again

Two events would change the shape of this page, and neither is announced.

The first is Muse Video shipping. A real release — with a resolution, a clip length and a price — would turn this from a status comparison into an actual head-to-head, and it would put Meta into a market where it currently has a leaderboard position and nothing else.

The second is an endpoint for Muse Realtime Avatar, or for the voice layer beneath it. Even a waitlist would move it from consumer feature to infrastructure decision. Until then, every architecture that includes an embodied Meta agent has one component nobody can implement.

There is a third possibility worth naming, because it is the one nobody plans for: Meta could ship Muse Video into the same app and let the two features meet. A prompt-to-video generator that produces a character, feeding a real-time renderer that animates it in conversation, is an obvious product. Neither Meta post suggests it is on the roadmap, and if it happens it will probably arrive the way the avatar did — a single post, no changelog, one country.