What can Meta AI do with music, and when do you need a dedicated tool?

You open Instagram, tap the Meta AI chat, and type: "make me a song for my sister, she's moving to Lisbon on Friday." What comes back is text. Maybe a decent verse, maybe a chorus that rhymes "Lisbon" with "missin'". No play button, no voice. So you search "meta ai music", and the first technical result is a GitHub repository that asks for Python 3.9 and PyTorch 2.1 before it will make a sound.

Both of those are "Meta AI music", and they have almost nothing in common. One is a chat assistant that lives inside apps you already use. The other is a family of research models for developers. A third thing, the music sticker on your Reels, is not generation at all, and a fourth, a song button inside Instagram DMs, has been reported in testing but is not documented on Meta's own help pages yet. This page separates them, says what each one does and does not do, and tells you where a dedicated tool earns its place.

One thing to get out of the way: sunomarket is an independent site. We are not Meta and not Suno, and we are not affiliated with either company or with anyone else named here. Everything below about Meta comes from Meta's own pages, linked at the points where it matters, and anything we could not confirm there is labelled as such.

Four different things share the name

Most of the confusion around "can Meta AI make music" comes from treating these as one product. They are built by different teams for different people.

If you came here because you want a finished song with someone singing your words, none of the four rows hands you that today in a documented, generally available way. That is the short answer. The rest of the page is about what each row is good for, and what to use instead for the parts it does not cover.

If the assistant you are actually weighing is Grok rather than Meta AI, the same question for xAI's chatbot is answered in can Grok make music. That page is for someone choosing between chat assistants. This one is for someone who already lives in WhatsApp and Instagram and wants to know how far Meta's own tools go.

What the Meta AI assistant does when you ask for a song

The assistant is a language model with a chat window. Since April 2026 it runs on a model Meta calls Muse Spark, and Meta's own announcement of that model talks about reasoning, reading what a camera sees, voice conversations and generating images. It does not mention generating music or songs. So when you ask the assistant for a song, it does the thing a language model does: it writes words.

That is not useless. For a birthday message, a toast, or a joke song for a group chat, a chat assistant gets you a first draft in seconds and costs nothing extra if you are already in the app. Where it gets thin is the stuff a song needs and a chat reply does not:

• Shape. A chat reply has whatever shape your message implied. Ask casually and you may get four stanzas with no chorus, or a chorus that changes its words every time it comes back.
• Section tags. Music tools that sing your text, Suno being the one we build for, read tags like [Verse] and [Chorus] to know which lines are which. A chat assistant adds them only if you ask, and keeps them consistent only if you keep asking.
• Singability. Long lines packed with adjectives read fine on a screen and fall apart once a melody has to carry them.
• Audio. There is none. Even a perfect lyric from the assistant still has to go into a separate tool that sings.

None of this is special to Meta. ChatGPT and Grok sit in the same spot, and the comparison of a general chatbot against a purpose-built lyric tool is worked through in can ChatGPT write good song lyrics. The practical rule is the same for all of them: a chat assistant is a fine notebook and a weak finisher.

The song option reported in Instagram DMs

In early October 2026, Social Media Today reported that Instagram had added a song option to the DM menu: you type lyrics, choose a genre, and get a generated track to send in the chat. The report is based on a creator's Threads post showing the feature, and quotes Meta as saying it could be a "fun engagement option".

Here is what we could not confirm on Meta's own pages at the time of writing: who has access, how long the tracks run, whether they carry a sung vocal, whether you can download the result, and what terms apply to it. Treat it as a test you may or may not see in your app. If it reaches you, it fits exactly one job well: a short, throwaway track for a friend inside a chat. If you need a file you can post elsewhere, edit, or keep, a DM toy is the wrong container, and that is before any question of quality.

Notice what the reported flow still asks you for: lyrics. Even Meta's own song button starts from words you bring. That detail matters for the rest of this page.

MusicGen and AudioCraft: real music models, built for research

This is the part of "Meta AI music" with real audio coming out of it. AudioCraft is Meta's open-source library for audio generation, announced in August 2023. Its GitHub repository lists the models it carries:

• MusicGen, text to music, optionally guided by a melody
• AudioGen, text to sound effects
• EnCodec, the neural codec underneath
• MAGNeT, a non-autoregressive text-to-music and text-to-sound model
• MusicGen Style, text plus a style reference
• JASCO, text to music conditioned on chords, melodies and drum tracks
• AudioSeal, audio watermarking, and Multi Band Diffusion, an alternative decoder

What Meta says about MusicGen on its own model card is the part most "Meta AI music generator" pages skip:

• It comes in 300M, 1.5B and 3.3B parameter sizes.
• It was trained on about 20,000 hours of licensed music from Meta's Sound Collection, Shutterstock and Pond5.
• "The model is not able to generate realistic vocals." Vocals were removed from the training data.
• It was trained on English descriptions and does worse in other languages.
• Its primary stated use is research on AI music generation.
• Code is released under MIT. Model weights are released under CC-BY-NC 4.0, a non-commercial licence.

A MusicGen request looks like this. There is no lyrics box, because there is nothing to sing them:

description: warm lo-fi hip hop, dusty Rhodes chords, soft vinyl crackle, slow head-nod drums
melody:      (optional) an audio file whose tune the output should follow
output:      a short instrumental clip; the model card's own example asks for 8 seconds

So MusicGen is the right tool for a narrow set of people: developers building something, researchers comparing models, hobbyists who want to learn how text-to-music works on their own GPU, and anyone who needs short instrumental beds for experiments. It is the wrong tool for a song with words, and the licence on the weights is a reason to read carefully before you put its output anywhere that earns money. How a non-commercial licence on weights applies to what the model produces is a question for a lawyer in your country, not for a blog post. This is not legal advice; our general position on AI music rights is in who owns AI-generated music.

If running a model locally is the appeal, MusicGen is not the only option, and some alternatives carry different licences and can produce vocals. They are compared side by side in open-source AI music generators you can run locally.

Movie Gen and SAM Audio: adjacent, not song tools

Two more Meta research projects show up in searches.

Movie Gen is Meta's video generation research. Its page says it can "create and extend sound effects, background music or entire soundtracks" that reflect the tone and rhythm of a video. That is music for picture: a bed under a clip, not a song with a verse and a chorus. The page does not say you can sign up and use it today, so we will not claim you can.

SAM Audio, released in December 2025, goes the other way: it separates sounds rather than making them. You point at a source with a text prompt, a click on a video frame, or a time span, and it pulls that sound out of the mix, across speech, music and general sound. Useful if you want to lift a vocal out of a recording or drop a noisy background. It will not write or sing anything.

Music in Reels and Stories is a library, not a generator

When people say "I added music with Meta" they usually mean the music sticker or the audio button in the Reels camera. That is a catalogue of existing recordings, licensed by Meta, with availability that varies by account type and country. Nothing is generated.

There is also Meta's own Sound Collection. Meta's developer documentation describes it as royalty-free tracks, and Meta offers them as replacements when a Reel with a detected copyright issue is about to be promoted as an ad. If your actual problem is "I need background music under a business Reel that will not get muted", Sound Collection is the tool to try first, inside Meta's apps, and you do not need anything from us.

Which job needs which tool

Here is the decision as one table. Pick the row that matches what you actually need to walk away with.

The row most readers land on is the third one, and it splits into two steps: the words, then the voice. Meta gives you a free notebook for the first step and nothing documented for the second.

Where a lyrics generator fits, and where it does not

The second step, the voice, is Suno's job or the job of a similar singing tool. We do not do it. Our AI song generator does not sing and never returns audio. It returns a title and a full lyric with [Verse], [Chorus] and other section tags, laid out to paste straight into Suno's Lyrics field. Suno is a separate company; we do not resell it, control it or speak for it.

What the generator changes compared with a chat prompt is that the brief has fixed fields, so the shape is decided before a single line is written:

• POV: first, second, third, collective or narrator
• Structure: classic, no-bridge, double-chorus, story, loop or ballad
• Hook: title-line, question, contrast, confession or chant
• Emotion: one of twelve, from love and longing to defiance and calm
• Vocal: unspecified, female or male, so the lines are written for the voice you plan to pick in Suno
• Language: nine, including English, Russian, Spanish, German, Japanese and Korean
• Age: up to three bands, or any, so the vocabulary fits the listener
• Theme: free text, up to 800 characters, where the sister, Lisbon and the Friday deadline go

Here is the Lisbon brief run through those fields, and the shape of lyric it is meant to produce. The lyric below was written for this page to show the shape; it is not a saved generator output.

POV: second | language: en | structure: no-bridge | hook: title-line
vocal: female | emotion: hope | age: 25-34
theme: my sister moves to Lisbon on Friday for a new job; last night in
the flat we shared; boxes in the hall; she wants to give back her key

Title: Keep the Spare Key

[Verse 1]
Tape on the boxes, your name in black marker
The kettle's the last thing we haven't packed
You hold out your key like it's something to borrow
I fold your hand shut and I push it back

[Chorus]
Keep the spare key, keep the spare key
Hang it where you'll see
Lisbon gets the rest of you
One door here still opens for you

[Verse 2]
You'll learn which tram climbs the hill to your office
You'll find your own bakery, your own café
And some Sunday night when the city feels foreign
You'll know there's a lock that still turns your way

[Chorus]
Keep the spare key, keep the spare key
Hang it where you'll see
Lisbon gets the rest of you
One door here still opens for you

Check it against the brief. Second person throughout: the singer talks to "you". No bridge: two verses and a chorus sung twice, nothing else. Title-line hook: the title opens every chorus. The theme's concrete details (the boxes, the key, Lisbon) are carried by the verses, which is what keeps a song like this from sounding like a greeting card. That fixed layout is what you pay for; a chat assistant can get there too, but you steer it there by hand, message after message.

What it costs, at the time of writing, with all prices in dollars:

• SM1, the base model: 20 credits per lyric, which is $0.20.
• SM2 and SM3, the more polished models: 80 credits, $0.80. Both run an extra editing pass in which a detector flags weak lines and a repair step rewrites them.
• Plans: $5 a month adds 500 credits (25 SM1 lyrics), and unspent plan credits are cleared at renewal. $19 once, the lifetime option, adds 1,500 credits once (75 SM1 lyrics) that never expire; it is not unlimited generation. A $10 pack of 1,000 credits can be bought from the profile on top of an active plan or lifetime, and those credits never expire either.
• No free trial of the lyric generator: new accounts start with 0 credits. What is free without a card is the style catalogue with full recipes for the ten most popular styles, all prompt guides, the first three preset categories of the prompt generator and the vocal gender toggle. Details are on pricing.

Our lyrics are non-exclusive: two people with similar briefs can get overlapping lines, and we do not promise a unique text or clear it against anyone else's rights. Commercial use is allowed under our terms, which are spelled out on the disclaimer page. Again, not legal advice.

When should you not pay for this? When the chat draft already says what you wanted. A dedicated generator earns its $0.20 when you are going to sing the result, where shape and tags decide whether Suno follows the song or wanders off it. For a line under a birthday photo, Meta AI is enough and costs you nothing extra.

Moving from a Meta AI draft to a finished song

If you already have words from the Meta AI chat and want them sung, the route is short:

1. Decide the shape first. Verse, chorus, verse, chorus is the safe default. Write it down before you edit a single line.
2. Make the chorus identical every time. Chat drafts often vary it. Pick the best version and paste it in every chorus slot.
3. Add tags. Put [Verse 1], [Chorus], [Verse 2] on their own lines. Inside the Lyrics box, anything not in square brackets gets sung, so stage directions belong in brackets too.
4. Shorten long lines. If you run out of breath reading a line aloud at normal speed, a singer will too.
5. Write the style separately. Genre, mood, instruments and voice go into Suno's Style field, not into the lyric. Our guide to Suno prompts covers what that field needs.
6. Generate a few takes in Suno and keep the one that sits best.

Or skip steps 1 to 4: put the same story into the generator's theme field, choose structure and hook from the lists, and the lyric comes back already tagged and in a fixed layout. Then go straight to step 5.

The quick version

Meta has a chat assistant that writes words, research models that make instrumental clips under a non-commercial weights licence, a sound separator, a music library for posts, and a reported song button in DMs. If what you need is a sung song with your own story in it, the plan is two tools: something that writes a tagged, singable lyric, and something that sings it.

Write tonight's version this way: take the story you were about to type into the Meta AI chat, put it in the theme field of the song generator, choose no-bridge and title-line, and paste the result into Suno with a style line of your own.