Magenta RealTime 2: Google's Free AI Music Model You Play Like an Instrument, Live on a Mac Today and Headed for PCs
The open music model built around latency, not song length, and what a new port changes for musicians without a Mac
Most AI music tools work like a vending machine. You type a description, you wait, a finished song drops out, and if the bridge is wrong you put in another coin. Nothing about that resembles playing music.
Magenta RealTime 2, an open model from Google DeepMind, is built on the opposite idea. It never hands you a finished song. It generates a continuous stream of instrumental audio and listens for your changes while it plays: a new text prompt, a recording whose style it should borrow, notes from a MIDI keyboard. Google's project page puts the gap between your move and the music answering at about 200 milliseconds, down from roughly three seconds in the first version. That is the difference between steering a band and leaving notes for one.
Until this week that instrument only played live on Apple Silicon Macs. On October 28 a community port appeared that claims to run it on ordinary PC graphics cards in about 2 GB of memory. Both halves of that story deserve a careful look, because one is ready to use and the other is not quite there.
Why response time is the whole product
A fifth of a second sounds like a spec sheet number. For a musician it is the line between a tool and an instrument.
When the music answers in two or three seconds, you cannot play with it. You make a change, wait, judge, make another change. It becomes a dialogue with a slow collaborator. When it answers in a fifth of a second, you can ride a fader, sweep between two styles, or push the groove somewhere new and hear the result while you are still moving. Google says each frame of audio is now 40 milliseconds long, against two seconds before, which is why the controls feel immediate.
The model also keeps a working memory of roughly the last 20 seconds of what it played, per the model card. That is enough for the music to stay coherent as you steer it, and short enough that it will follow you somewhere new rather than clinging to the opening idea.
The trade for all this speed is scope. Magenta RealTime 2 makes instrumental music. The card says the training data is mostly instrumental, and while specific prompts can coax out vocal sounds, they tend to be wordless. If you need a sung chorus, this is the wrong tool. If you need a bed under a live set, a soundtrack that reacts to a performer, or a sketchpad for grooves, it is one of the most interesting open options anyone has shipped.
What you can steer, and how
You have three ways to push the music around.
Words. Describe a style ("disco funk," "slow ambient pads with tape hiss") and the model aims there.
A recording. Feed it an audio clip and it borrows the style. It does not continue that recording or copy its melody. Think "make it feel like this," not "pick up where this left off."
Notes. A MIDI keyboard can tell the model which pitches to sound and when, so you can hold a chord and let it fill in the arrangement around you.
Google ships the model with free tools that turn those controls into something playable, listed on the MRT2 page:
- Jam, a standalone app for playing along with style presets and MIDI control.
- Collider, which lays prompts out on a 2D surface so you can slide between them and land on blends that have no genre name.
- MRT2, a standalone app plus an Audio Unit plugin that drops onto a MIDI track in a Mac DAW such as Logic.
- Extensions for Max/MSP, Pure Data and SuperCollider, for people who build their own performance rigs.
The page lists no price. You download a zip, and the apps fetch the model on first launch: about 450 MB for the Small version and about 2.5 GB for Base.
Which Mac you need
Two sizes exist. Small is 230 million settings learned in training; Base is about ten times bigger at 2.4 billion. Bigger usually means richer output, at the cost of needing more machine.
Per Google's page, Small plays in real time on any Apple Silicon MacBook, including a MacBook Air. Base needs an M3 Pro or M2 Max class chip or better to keep up live. Both will generate offline, slower than real time, on any Apple Silicon Mac. Google's Python library can also run offline on NVIDIA graphics cards, per the project README, but the live, playable experience has been Mac only.
That is the gap the new port tries to close.
The PC port: promising numbers, early days
The port is audio-cpp/Magenta-RealTime-2-GGUF, created on October 28 according to Hugging Face. It repackages Google's model for audio.cpp, a free local engine that already runs speech, music and voice models on Windows, Linux and Mac, on NVIDIA cards, on AMD and Intel cards through a cross-vendor graphics standard called Vulkan, and on plain processors.
The numbers on the card are striking. On a single RTX 5090, the porter's speed figures work out to the Small model making a 30-second clip in roughly two and a half seconds, using about 2,150 MiB of graphics memory, roughly 2 GB. Google's own Python code, on the same card, peaked at about 24 GiB. Base needs about 4.1 to 6.2 GiB depending on the file version, which puts it within reach of a mid-range gaming card on paper. In streaming mode the engine hands back audio in blocks of up to 160 milliseconds, and the first sound arrives in about a quarter of a second.
If those figures hold up on cheaper hardware, a musician with a gaming PC could run the same model a MacBook Air runs. That is the reason this port is worth watching.
Now the caveats, because they are real.
These are the porter's own numbers. One top-end graphics card and four requests per setting. Your laptop will be slower. How much slower nobody has published yet.
The engine may not be ready for it. The card's commands assume audio.cpp knows about Magenta. But the newest audio.cpp release, v0.9.0 from September 30, does not mention it, and the model guide the card links to would not load when I tried it on October 28. My guess is that support sits in the engine's unreleased code and has not reached a release you can download and double-click. If the commands fail on the current release, that is probably why.
It is a command line, not an instrument. The port gives you a terminal command that writes a file or streams blocks of audio. There is no Jam, no Collider, no plugin. The card also says each request starts fresh music, so it does not yet keep one continuous performance going across your changes the way Google's apps do. Somebody will build a playable front end on top of it. Nobody has yet.
So on a PC today, think of this as a fast way to generate instrumental sketches and test what the model sounds like on your hardware. The live instrument still lives on the Mac.
What the licence lets you do
This is where Magenta RealTime 2 compares well with most open music models.
The model files carry CC-BY 4.0, a licence that allows commercial use as long as you credit the source. The code is Apache 2.0, also commercial friendly. Google adds usage terms, and the important line for creators is that you must not generate outputs that infringe someone else's rights. The card also says "Google claims no rights in outputs you generate," and puts responsibility for those outputs on you.
In plain terms: you can use what you make in paid work, Google does not claim it, and if you prompt it toward someone's recognizable track, that is your problem. The port inherits the same terms because it is a repackaging, not a new model.
Compare that with YuE2, another popular open music model, whose weights are non-commercial. For a working composer, that difference decides which tool you can bill for.
Put this into practice
On an Apple Silicon Mac:
- Download the bundle from the MRT2 page and unzip it.
- Set your audio interface or DAW to 48 kHz first. Google's page asks for this, and a mismatch is the first thing to rule out if something sounds wrong.
- Start with Jam and the Small model, even if your Mac can run Base. Small loads faster and lets you learn the controls without stutter.
- Pick two style prompts that should not go together and sweep between them. This is where the model earns its keep: the in-between states are often more useful than either end.
- Plug in a MIDI keyboard, hold a chord, and listen to how the arrangement bends around it.
- When you have something you like, move to the MRT2 plugin on a MIDI track so you can record it into a session.
On a PC with an NVIDIA, AMD or Intel graphics card:
- Install audio.cpp from its releases page.
- Download the Small file,
magenta-rt2-small-f16.gguf, from the port's card. - Run the card's ten-second test command with
--backend cudaon NVIDIA or--backend vulkanon AMD and Intel. - If it fails, check the audio.cpp project for a newer release before you troubleshoot your setup. Support may simply not have shipped yet.
- If it works, time a 30-second clip. Faster than 30 seconds means your card could keep up live once a playable front end exists.
Where it falls short
No words. Instrumental only, by design and by training data.
No published scores. The card says evaluation results will come in a technical report that has not appeared yet. Everything you know about quality, you will learn by listening.
Base needs a serious Mac. The bigger model plays live only on higher-end Apple chips.
The PC path is unfinished. Speed figures from one top-end card, runtime support not in a published release, and no interface for live play.
Style borrowing is loose. Feeding it a recording copies a feel, not a part. If you wanted it to continue your demo, it will not.
The instrument is the interesting part
AI music has spent two years racing toward longer, more polished finished songs. Magenta RealTime 2 bets on something else: that musicians want a responsive partner they can push around, more than a jukebox that takes requests.
That bet is easy to test if you have a Mac. Open Jam, give it ten minutes, and notice whether you start playing it or just listening to it. If you are on a PC, run the port's test command and watch the clock. Either way you will know quickly whether this belongs in your rig, and nobody's benchmark can tell you that.