Meta has described what Muse Video does at the level of capability and published almost nothing at the level of specification. That gap is where the confident blog posts live. This page separates what has been stated from what follows from the model, and names the specs length, resolution, formats, rights that nobody outside Meta actually knows.
These are stated at product level rather than measured, which is exactly why they are durable: none of them is a number that a rollout wave or a plan change can quietly alter. The tag on each row says whether Meta said it or whether it follows from something Meta said.
The core of it: you write what you want and it generates video rather than storyboards or stills. Short is a property of the medium, not a setting someone chose why coherence caps the length.
Audio is part of what the model generates, not a track laid on afterwards. That makes it a dimension to judge on every attempt, and a reason a clip can be unusable for a reason you cannot re-edit. Audio in detail.
Video runs on Muse Spark 1.3, released September 2, 2026, which accepts text, images and video as input. One model behind Meta AI, Muse Image and video is the reason capabilities arrive in step.
No separate download and no separate account. If the Mac app is installed and the feature has reached you, video is in the same chat as everything else on a Mac specifically.
Of everything that meters on the free tier, video bites first, and Meta One raises the ceiling rather than removing it. The figures themselves are unpublished what we can say about cost.
Video has its own rollout, narrower than Meta AI’s, and being able to chat says nothing about being able to generate. The five access gates.
“Accepts images as input” is a fact about the model. It is not a promise that the video box in your copy of the Mac app, in your country, on today’s build, will let you attach one. Model capabilities and exposed features are different layers, and staged rollouts pull them apart.
The interface is the authority on what you can do; the model spec only tells you what is possible in principle. Where this page marks a row as following from something rather than stated, treat it as a hypothesis to test in the product.
Muse Spark 1.3 arrived on September 2, 2026 with a one-million-token context window and multimodal input text, images and video. For a chat assistant that reads as a bigger memory. For video generation it reads as something more concrete: a clip request can be conditioned on things other than words.
A reference frame is worth more than a paragraph of adjectives, because style is far easier to show than to describe. That is an argument for settling the look in cheap stills first, whatever the video surface currently accepts.
The shared engine also explains the release pattern people find confusing. Model improvements land underneath all three products at once, so video quality can shift without a video announcement and a model that got better at tool use, as 1.3 did, is not necessarily better at holding a clip together.
Worth keeping in proportion: 1.3 was not a clean sweep. Independent testing found long-document accuracy slipped from 83% to 79% against 1.2, which is a reminder that a version number going up is not a guarantee on every axis including the ones that matter to a clip.
Each row here is a question we are asked and cannot answer honestly. What we can give you is the reason it is unpublished and the move that works anyway which is usually cheaper than the number would have been.
| The question | Why there is no figure | What to do instead |
|---|---|---|
| How long can a clip be? | Unpublished, and in practice a moving target coherence, capacity and plan all bear on it. | Write for a few seconds. Treat length as a budget, not a setting. |
| What resolution and frame rate? | Never stated, and likely to differ by surface and plan. | Generate one clip and inspect the actual file before you plan an edit around it. |
| Which aspect ratios? | Not documented; vertical and square are the obvious targets but obvious is not confirmed. | Ask for the framing in the prompt, then check what arrives social formats. |
| What file do I get, and can I download it? | Export behaviour is a product detail Meta has not spelled out and can change between builds. | Test the whole path generate, save, open in your editor before committing to a deadline. |
| How many clips per day or week? | Quotas on a staged feature vary by country, plan and rollout wave; a single figure would be wrong for most readers. | Watch your own limit messages. Tiers raise the ceiling. |
| How long does generation take? | Depends on load, region and what you asked for. Any quoted time is one person’s afternoon. | Assume you will run several attempts, and cost the review time in. |
| Can I use clips commercially? | A terms question, not a capability, and the answer can differ by market. | Read Meta’s terms directly; see also what Meta says it uses. |
| Is there an API for video? | Nothing announced for video specifically. | Track the Meta Model API page rather than assuming parity. |
| Which languages does the audio speak? | Unstated, and language coverage rarely matches the feature’s country coverage. | Test your language early rather than at the end of a project. Audio notes. |
If you find a specification quoted with confidence elsewhere, check whether it is sourced to Meta or to one person’s test. Both are useful; only one is a specification.
Not oversights, and not coming in a point release. Each one is ruled out by how the thing works, which makes them safer to plan around than any published limit.
There is no editor because there is nothing to edit into: frames are produced as one coherent run rather than assembled from independent pictures, so a bad half-second cannot be replaced the way a bad object in an image can.
Plan around it: decide the shot before generating, and do any real cutting in a normal editor afterwards, on the file you were given.
Asking for more time is asking for more coherence, and a new run will differ in places you never mentioned. Even where a continuation exists as an interface button, it is a fresh generation conditioned on the old one not an append.
Plan around it: if the idea needs twenty seconds, plan several short clips with a cut between them rather than one long one.
Generation is not deterministic, so the clip you liked yesterday is not retrievable by re-running the words. There is no seed to write down that makes it come back.
Plan around it: keep every attempt you might want. Deleting a near-miss is deleting the only copy. Working habits for creators.
These follow from how generative video works rather than from a Meta statement about Muse Video specifically the same constraints show up in competing tools.
Professionals do this all the time with tools that publish nothing. The trick is to establish the facts once, for your account, and then design inside them.
Generate something you do not need, then look at the file properly: duration, dimensions, sound, what the download gives you. Fifteen minutes of this replaces every spec you cannot find. Where to start.
One idea, one movement, no plot. If the concept collapses when the clip is short, the concept not the tool is the thing to change. Prompts that survive.
Style is cheap to iterate as images and expensive to iterate as clips. Lock the palette and the subject in Muse Image, then describe that look to the video model.
Whatever you generate has to land somewhere with its own rules on dimensions, length and AI disclosure. Confirm what the destination demands before you spend an allowance discovering it. Social destinations.
Plan on several generations plus watching each one through to the end, and keep a stills fallback for the version where video does not come good in time. When a still will carry it.
Model changes land under all three products at once and without a video announcement. Your own throwaway clip is the only reliable changelog. What changed recently.
At capability level: it turns a description into a short clip, it produces audio as part of generation rather than as a later step, it runs on Muse Spark 1.3, it lives inside the Meta AI surfaces you already have, it meters against your usage allowance more heavily than anything else, and it rolls out country by country. What has not been announced is numbers the six in full.
Meta has not published a figure, and any specific number you see quoted is someone’s guess or an observation from one account on one day. Treat clip length as short and variable, and plan the shot so that a few seconds is enough.
Not published. This matters most if you are cutting Muse Video material into footage from another source, so the safe approach is to generate one clip, inspect the file you actually get, and build the edit around that rather than around an expectation.
Yes, and the important part is that audio is part of what it produces rather than something added afterwards. That makes sound another dimension of the result to judge, and another reason a clip can be wrong in a way no re-edit will fix audio in detail.
Muse Spark 1.3 accepts text, images and video as input, so image-conditioned generation is architecturally available to the family. Whether the video surface exposes it to you, in your country, on your build is a separate question check the interface rather than the model spec.
No frame-level editing exists, because frames are generated as one coherent run rather than assembled from independent pictures. Changing anything means generating again, and the next attempt will differ in places you never mentioned the three absences.
Expect labelling and provenance signals on synthetic video, since it is the output class carrying the heaviest disclosure obligations, and those obligations vary by jurisdiction. The exact markings applied to your file are Meta’s to document and worth checking before you publish anywhere it matters.
That is a terms question rather than a features question, and it is the one worth reading properly rather than inferring. Commercial use, ownership and platform disclosure rules all sit in Meta’s terms, which supersede any guide including this one.
Nothing has been announced for video specifically. Meta’s developer surface is tracked separately on our Meta Model API page, and a model being available to developers does not imply the video feature is.
Because invented specifications are worse than none. Limits on a staged generative feature differ by country, plan and rollout wave, and a confident wrong number costs you a ruined edit rather than an inconvenience what is genuinely unpublished.
Duration, dimensions, sound, and what the download actually gives you fifteen minutes with a clip you will throw away establishes all of it for your account, which is the only account that matters.
Independent guide not affiliated with Meta. Clip lengths, resolutions, formats, quotas and commercial terms are Meta’s to state; confirm them in the product and in Meta’s own documentation before planning around them.